paid work for women and domestic violence: evidence from

Paid Work for Women and Domestic Violence:Evidence from the Rwandan Coffee Mills

Deniz Sanin§

January 17, 2022Click here for the most recent version

Abstract

Using the government-induced rapid expansion of the coffee mills in Rwanda in the 2000s, I first provide

causal evidence that a mill opening increases women’s paid employment, women’s and their husbands’

earnings and decreases domestic violence. Then I provide evidence which suggests that the decline in

violence is driven by women’s paid employment. The increase in husbands’ earnings is not the dominant

mechanism. The opening of a mill affects coffee farmers who reside in its catchment area, a buffer zone

around the mill, during the harvest months. A mill opening enables women to transition from being

unpaid family workers in their family plots to wage workers in the mills. Using a staggered difference-

in-differences (DID) design, I show that upon a mill opening, women in the catchment areas are 18%

more likely to work for cash and 26% less likely to self-report domestic violence in the past 12 months.

Using novel monthly administrative records on the universe of hospitalizations for domestic violence and

a DID event-study design, I also show that it is 20% less likely for the hospitals in the catchment areas

to have a domestic violence patient in a harvest month compared to one month before the beginning

of the harvest season. The decline in violence is present even among couples where husbands work

in occupations with no change in earnings with a mill, non-agricultural manual jobs. An increase in

women’s outside options and their contribution to household earnings, not exposure reduction between

couples, explain the results.

Keywords: Domestic violence, female labor supply, agriculture, Rwanda

JEL Codes: J12, J16, J21, J31, O12

§Economics Ph.D. Candidate, Georgetown University, Department of Economics & Pre-Doctoral Research Fel-low, Women and Public Policy Program, Harvard Kennedy School, [email protected]. I am extremely gratefulto my advisor Garance Genicot for her invaluable guidance and support on this project. I am very grateful to mycommittee members Laurent Bouton, Mary Ann Bronson and Martin Ravallion for their detailed comments and sup-port. I owe special thanks to Andrew Zeitlin and Leodimir Mfura, who helped me in my data application process.For very helpful comments, I would like to thank Daron Acemoglu, Marcella Alsan, Sonia Bhalotra, Alberto Bisin,Dara Kay Cohen, Carolina Concha-Arriagada, Jishnu Das, Leopoldo Fergusson, Claudia Goldin, Rema Hanna, BillyJack, Larry Katz, Roger Lagunoff, Luis Martinez, Ameet Morjaria, Nathan Nunn, Alex Poirier, Natalia Rigol, DaniRodrik, Frank Schilbach and Daniel Valderrama. This project has human subjects approval from Rwanda NationalEthics Committee. I acknowledge the financial support of Georgetown Graduate School’s Dissertation Travel Grantand GradGov Research Project Award.

https://denizsanin.github.io/MyWebsite/CoffeeJMP_DenizSanin.pdf

mailto:[email protected]

1 Introduction

Domestic violence is an extreme form of gender inequality, violation of human rights and a globalhealth problem of epidemic proportions (WHO, 2013). About 1 in every 3 women worldwidehave experienced either physical and/or sexual violence from their partners in their lifetime (WorldBank, 2015). To address domestic violence, providing employment opportunities to women isdiscussed in the policy debate. The argument is that employment can decrease violence via an in-crease in women’s outside options and bargaining power and/or a reduction in the financial stressin the household. Yet, causal evidence on the effects of increased job availability for women on do-mestic violence is limited. The domestic violence literature focuses on the effects of cash-transfers,unemployment, gender wage gap, education and dowry payments, which are different than havinga job.1 Moreover, evidence from cash transfers suggest that an increase in women’s resources mayincrease as well as decrease violence due to husbands’ incentives to extract the resources or malebacklash (Angelucci 2008, Bobonis et al. 2013, Hidrobo et al. 2016, Haushofer et al. 2019). Doesproviding jobs to women make them less or more vulnerable to domestic violence?

This paper investigates whether providing employment opportunities to women decrease theviolence they face from their partners using the government-induced rapid expansion of the coffeemills in Rwanda in the 2000s as a natural experiment. The expansion is an ideal setting to study theeffects of increased job availability for women since a mill opening enables women to transitionfrom being unpaid family workers in their family plots to wage workers in the mills for the sametasks as before. Thus, the results capture the effects of having a paid job and moreover are notinfluenced by learning a new skill.2

In 2002, the Rwandan government adopted the National Coffee Strategy that aimed to shiftto mill-processed coffee production to participate in the international specialty coffee market(Boudreaux, 2011). In the early 2000s, a public-private partnership project helped farmers toestablish cooperatives and build mills in their communities. After the project, farmers continued tobuild mills across the country. From 2002 to 2012, the number of mills increased from 5 to 213.3

A coffee mill is where coffee cherries, harvest of the coffee tree, are processed into coffeebeans for export. The context provides two key features for identification. First, a mill openingenables women’s transition to paid employment and creates a time-variation. Before a mill open-ing, women process coffee at home as a female-dominated task. Their husbands sell the home-

1See Section 2.2Having a job is a combination of earning income and non-monetary benefits of employment including social

network and psychosocial benefits (Hussam et al., 2021).3In 2012, coffee accounted for nearly 20% of Rwanda’s exports and 15% of its GDP (Macchiavello and Morjaria,

2020).

1

processed coffee in the local market as a male-dominated task and receive the income from coffeeas their personal earnings. After a mill opening, husbands within the catchment area sell coffeecherries to the mill. The mill demands paid labor for the coffee processing tasks during the harvestmonths, March-July, within a year. Since the tasks are female-dominated, a mill provides the op-portunity for women in the catchment areas to transition from unpaid family workers to daily wageworkers in the mill and have personal earnings. Given that the husbands were already generatingincome, there is no change in their type of earnings. Yet, husbands earn a higher income from sell-ing cherries to mills compared to selling home-processed coffee in the local market (Macchiavelloand Morjaria, 2020). Second, the opening of a mill creates a spatial variation. A mill serves coffeefarmers that reside within its catchment area, approximately a 4 km radius buffer zone around themill.4 It has a specific catchment area since cherries will rot if coffee farmers do not transport theircherries to a mill within few hours of harvest.

To causally identify the impact of a mill opening, I first use a staggered DID design exploitingthe spatial variation, within-outside of the catchment area, and the yearly time variation, before-after a mill opening. For this strategy, I use data on self-reported domestic violence and labormarket outcomes. Then I complement this analysis using novel monthly administrative records,the universe of hospitalizations for domestic violence, during the end of the expansion where thenumber of mills is fixed. I use a DID event study design exploiting the same spatial variation asbefore and monthly time variation within a year. I use the beginning of the harvest season, March,as the event where the mills start to operate.

Rwandan Demographic Health Surveys (DHS) provide information on couples’ self-reportedlabor market outcomes and women’s domestic violence for the past 12 months. Hospital Man-agement Information System (HMIS) data provides the universe of monthly hospitalizations dueto domestic violence. I combine data on the universe of mills using Rwanda GeoPortal and Mac-chiavello and Morjaria (2020). All aforementioned datasets are geocoded that enables me to linkthe couples and hospitals with the mills based on the GPS coordinates. I also use an additionalhousehold survey that provides couples’ individual earnings.

Upon a mill opening, being exposed to a mill increases the probability of working for cash inthe past 12 months by 18% with respect to the sample mean (0.40). The probability of working inthe past 12 months remains unchanged. Importantly, being exposed to a mill decreases the proba-bility of self-reporting a domestic violence experience in the past 12 months by 26% with respectto the sample mean (0.35). There is no statistically significant change in husbands’ probability ofworking and type of earnings being cash in the past 12 months. The occupation for each spouseremains unchanged. To establish these results, I use couples who reside outside of the catchment

4Farmers outside of the catchment areas continue to process coffee at home after a mill opening.

2

area but are located within the same district with the mills as the control group. The results arerobust to using another control group, the couples who reside within the donut area between 4 and8 km from the mills. Using the additional household survey data, I also show that upon a millopening, mill exposure increases both women’s and their husbands’ last daily earnings.

Next, using monthly hospitalizations data in a DID event study design, I test for an effectduring the harvest months, the only period mills operate within a year. Within the harvest months,May-July is the peak of the harvest where majority of the community around the mills work in themills. I show that it is 18% and 20% less likely for hospitals in the catchment areas to a have adomestic violence patient in June and July respectively compared to one month before the mills’month of operation, February. There are no statistically significant changes both for January, twomonths before mills’ month of operation, as well as for the post-harvest months when mills donot operate and thus women do not work for pay. As a placebo test, I present results for women’shospitalizations other than domestic violence. I find no changes. This rules out the concern thatwomen go to a hospital less during the harvest season due to increased opportunity cost of time.Moreover, finding a decline in domestic violence both with administrative and self-reported datasuggests that the results are not subject to reporting bias.

In the second part of the paper, I investigate the mechanisms behind the decline in domesticviolence. I first present a conceptual framework. The husband decides whether to inflict violenceand the wife decides whether to separate. The level of satisfaction with the marriage is privateinformation for each spouse. A mill opening affects violence in three ways. First, as the wifetransitions to paid work, her outside option, the utility of being separated, improves. This increasesthe probability of her initiating separation. Opposingly, the husband is incentivized to chooseviolence to extract the newly acquired resources from the wife. Third, as the household earningsincreases either via an increase in his or the wife’s earnings, husband’s marginal non-monetarybenefit from violence (e.g. financial stress relief) decreases. When the wife’s outside option due toher job is good enough, the husband’s extraction rate and his non-monetary utility from violencebased on household earnings is small enough, violence decreases.

To uncover the mechanisms behind the decline in violence empirically, I first show that upon amill opening, women in the catchment areas are more likely to make household decisions jointlywith their partners including large household purchases and contraception usage. This suggeststhat a mill opening improves women’s outside options and increases their bargaining power in thehousehold. Second, I show that the decline in domestic violence is observed even among cou-ples where husbands work in occupations with no change in earnings with a mill, non-agriculturalmanual job.5 The magnitude of the decline in violence is also similar to the result based on the

5Plumber, construction worker etc. Occupations do not change with a mill opening.

3

whole sample. This suggests that the increase in husbands’ earnings is not the dominant mech-anism behind my results. The increase in women’s outside options and decrease in husbands’non-monetary benefit from violence (e.g. financial stress relief) due to women’s contribution tohousehold earnings are plausible mechanisms. Moreover, unlike farmer couples, these couplesdo not work together. A mill opening is not a shock to the time they are exposed to each otherduring work hours. Thus, exposure reduction between couples is also ruled out as the dominantmechanism.6

All plausible channels are due to women’s paid employment. Thus, the paper suggests thatwomen’s paid employment is the driver behind the decline in domestic violence. The seasonalityof the decline in hospitalizations supports this and also speaks to the mechanisms. The resultsuggests that in a context where wages and household resources are low, access to credit marketsand saving technologies are limited, women having better outside options and husbands having lessnon-monetary benefit from violence based on household earnings is concentrated when womenhave the jobs.

I also investigate the dynamic impact of a mill opening using a DID event study specificationexploiting the number of years couples are exposed to a mill opening. Dynamic estimates showthat the effects on domestic violence persists for 4 years. After 4 years, there is no decline indomestic violence although women continue to work for pay. This suggests that the effects ofwomen’s paid employment in low-paid jobs may not persist for long years. For the years before amill opening, the coefficients are close to zero and statistically insignificant for all variables.

Recent econometric literature on DID estimators that use variation in treatment timing raisesconcerns about the validity of estimation results in the presence of treatment effect heterogeneity.My results are robust to using estimators proposed in de Chaisemartin and D’Haultfœuille (2020)and Sun and Abraham (2020) that gives valid results even if the treatment effect is heterogeneousover time and across groups. I also perform robustness checks related to catchment area measure-ment. I use different catchment area sizes, 5 and 10 km. I show that as the buffer radius increases,the more untreated couples are counted as treated, the effects of mill exposure fade out for all ofmy main outcome variables. This rules out the concern that the couples residing right outside ofthe catchment area are sorting themselves into the treatment group.

The organization of the paper is as follows. Section 2 reviews the related literature and high-light the contributions of the paper. Section 3 provides background information on the Rwandancoffee industry, rapid expansion of coffee mills and women’s paid employment in the coffee value

6Exposure reduction mechanism argues that there is less violence when the wife work in the mills simply becausethe couple interact less frequently. I do not model it in my conceptual framework, however, it is documented as achannel in the literature.

4

chain. Section 4 introduces all data sources. Section 5 outlines the identifying assumptions, threatsto identification and propose the identification strategy. Section 6 present the empirical results,Section 7 introduces the conceptual framework that guides the mechanisms behind the results.Section 8 uncover the mechanisms empirically. Section 9 provides robustness checks. Section 10concludes.

2 Related Literature

This paper contributes to several strands of the literature. First, it contributes to the literatureon women’s income and domestic violence. The literature provides evidence for both a positiveand negative relationship. The negative relationship between the two is explained with the theoryon the increase in the outside option, the decline in financial stress and exposure reduction. Thepositive relationship is explained with instrumental/extractive theory of domestic violence andmale backlash. In the theory on the increase in the outside option, decline in financial stress andthe instrumental theory of domestic violence, violence is incorporated into household bargainingmodels via men’s motives.7 If a man is violent because it contributes to his utility directly, via arelease of stress or frustration, then the violence is expressive (Gelles 1974, Tauchen et al. 1991).If a man is violent to extract resources from his wife to increase his consumption of goods or tocontrol the wife’s behavior, then the violence is instrumental (Gelles 1974, Tauchen et al. 1991).I briefly review the theories and the literature on the relationship between women’s income anddomestic violence below.

I. Increase in the Outside Option. An economic theory of household bargaining that incor-porates only expressive violence (Farmer and Tiefenthaler 1997, Aizer 2010, Anderberg et al.2016, Hidrobo et al. 2016) or expressive and instrumental violence combined (Tauchen et al. 1991,Haushofer et al. 2019), suggests that an increase in women’s income decreases domestic violenceby improving her outside option and thus her bargaining power. For the choice of outside options,the models use either value of being divorced as in the divorce threat model (Manser and Brown1980. McElroy and Horney 1981) or noncooperative equilibrium within the marriage as in theseparate spheres model (Lundberg and Pollak, 1993)

II. Expressive Violence (Effect of Income). In the literature, expressive violence is used to studythe effect of income on domestic violence. Tauchen et al. (1991) highlights that expressive vio-lence does not necessarily need to be sadistic but may allow the husband to release economic stress.Similarly, Angelucci (2008) argues that an increase in spousal income may reduce the husband’s

7See Chiappori and Mazzocco (2017) for a detailed review of household decisions.

5

economic stress and decrease domestic violence. Haushofer et al. (2019) incorporates expressiveviolence into their model to analyze the relationship between economic stress and domestic vio-lence. The paper highlights that if violence and the husband’s wealth are substitutes in his utility’sexpressive component, then they are also substitutes in his total utility. Thus, an increase in thehusband’s income decreases the violence he inflicts on his wife. The theory is in line with evidencefrom behavioral neuroscience that finds a link between stress and aggressive behavior (Kruk et al.,2004).

III. Exposure Reduction. Criminologists develop the theory of exposure reduction to argue thatan increase in female employment may reduce domestic violence due to the decline in the timecouples spent together (Dugan et al., 1999).

IV. Instrumental Violence (Resource Extraction). There also exists theories which suggest thatthere is a positive relationship between women’s resources and domestic violence. Bloch and Rao(2002), Eswaran and Malhotra (2011), Bobonis et al. (2013), Anderson and Genicot (2015) andCalvi and Keskar (2021) present bargaining models that incorporates instrumental violence wherethe husband uses violence or threats of violence as a bargaining tool to extract resources from hiswife or to enhance his bargaining power. These models argue that violence is an instrument inhousehold bargaining for men and an increase in women’s resources also increases the domesticviolence they face.

V. Male Backlash. Theory of male backlash proposed by sociologists also argues that an increasein women’s income may indeed increase domestic violence. According to the theory, female em-ployment challenges the husband being the “breadwinner” and thus, he inflicts violence on his wifeto reinstate his traditional gender role (Macmillan and Gartner, 1999). Although both instrumen-tal violence and male backlash theory suggest a positive relationship between female employmentand domestic violence, male backlash theory do not necessarily argue that the husband’s motive toinflict violence is to extract resources from the wife. The theory focuses on the husband perceivingthe wife’s employment as a status threat in the household due to entrenched social norms.

A growing empirical literature on domestic violence provide evidence for the theory on theincrease in women’s outside options (Aizer 2010, Anderberg et al. 2016, Hidrobo et al. 2016,Haushofer et al. 2019), reduction in financial stress (Angelucci 2008, Bhalotra et al. 2019, Heathet al. 2020, Arenas-Arroyo et al. 2021, Bhalotra et al. 2021), exposure reduction (Chin, 2011),instrumental violence (Bloch and Rao 2002, Eswaran and Malhotra 2011, Bobonis et al. 2013,Heath 2014, Anderson and Genicot 2015, Erten and Keskin 2018, Bhalotra et al. 2019, Erten andKeskin 2018, Erten and Keskin 2020, Calvi and Keskar 2021 Erten and Keskin 2021) and malebacklash (Angelucci 2008, Luke and Munshi 2011, Tur-Prats 2021, Alesina et al. 2020, Guarnieriand Rainer 2021). It is worth mentioning that within the empirical literature on domestic violence,

6

La Mattina (2017), Rogall (2021) and Sanin (2021) focus on domestic violence in the Rwandancontext, yet they do not investigate the relationship between women’s resources and domestic vio-lence. They focus on the long-run impact of the Rwandan Genocide, female political participationand the effects of domestic violence laws on domestic violence in Rwanda respectively.

This paper also contributes to the literature on female labor force participation (FLFP) in devel-oping countries. Since FLFP is strikingly low in the developing world (non Sub-Saharan Africancountries), the literature focuses on the drivers of FLFP and the effects of interventions that canfoster FLFP. These include cultural practices, trade liberalization, residence patterns, microcredit,education, cash-transfers, correcting husbands’ beliefs, laws, opening bank accounts for womenand psychosocial interventions (Alesina et al. 2013, AlAzzawi 2014, Gaddis and Pieters 2017,Khanna and Pandey 2020, Angelucci et al. 2015, Banerjee et al. 2015, Keats 2018, Baird et al.2019, Bursztyn et al. 2020, Hyland et al. 2020, Field et al. 2021, McKelway 2021). In this paper, Ifocus on the consequences of women’s employment, specifically paid employment. Heath and Jay-achandran (2017) reviews the litearature on FLFP in the developing countries and highlights thatincreased job availability for women is found to affect women’s marriage, fertility and educationdecisions (Jensen 2012, Sivasankaran 2014, Heath and Mobarak 2015), children’s health (Qian,2008) and bargaining power (Anderson and Eswaran 2009, Majlesi 2016). My results suggestthat the interventions that increase female paid employment have the potential to benefit womenbeyond their labor supply and decrease the violence they face from their partners.

3 Institutional Context

In this section, I provide background information on the coffee industry in Rwanda, rapid expan-sion of the mills following the agricultural reforms in the early 2000s and women’s work opportu-nities in the mills.

3.1 The Coffee Industry in Rwanda

Coffee production and processing. The first two stages of the coffee value chain, cultivation andprocessing are the stages of interest throughout the paper. Both stages includes labor intensiveprocesses which affect the quality of the end product.

A coffee tree produces coffee cherries that contains coffee beans. It takes at least three to fiveyears for a coffee tree to produce cherries after it is planted. When the cherries ripen, they shouldbe harvested (picked) by hand. The cherries do not ripen all at once which makes harvesting a

7

labor-intensive process. The harvest season lasts for approximately four to five months. It lastsfrom March to July in Rwanda and the peak of the harvest is from May to July.

After harvest, coffee needs to be processed. Processing is made up of three main tasks. Thefirst task is cleaning, which is removing the outer layer (pulp) of the coffee beans. In the secondtask, the coffee beans are dried under the sun on flat surfaces.8 In the third task, the defective driedcoffee beans are sorted out by hand based on their color and size, which is a very labor-intensiveprocess like harvesting.

There are two possible methods to process the coffee cherries: dry processing and wet process-ing. In the dry processing, which is also known as the traditional method due to being the oldestcoffee processing method, all tasks are done by farmers at their homes without any machinery.9

The outcome of this process yields a low quality product which is sold in the local market for avery low price.

In the wet processing method, farmers sell their coffee cherries to coffee mills (called washingstations in Rwanda). A mill can be thought as a large firm for developing country standards wherecoffee is wet-processed. In the mills, cleaning is done with specific machinery which uses plentyof water. That is why the method is called wet processing. Drying and sorting tasks are done byhand as in dry processing. Mills demand seasonal wage labor from the neighboring community forthese tasks. The method yields a high quality product which is sold in the international market fora high price to international buyers.10 Multinational companies like Starbucks is an example of aninternational buyer of Rwandan wet-processed coffee.

3.2 Rapid expansion of the mills

Agricultural Reforms. In 2000, Paul Kagame came into power and prioritized economic growthto rebuild the country in the aftermath of the Rwandan Genocide (1994). He launched the Vi-sion 2020 program in 2000 (Boudreaux, 2011). The program outlined a list of goals which thegovernment aimed to achieve by 2020. One of the main goals was to transform agriculture intoa high value sector. In light of this goal, the government adopted the National Coffee Strategy in2002 which aimed to shift to high quality, wet-processed coffee production to participate in theinternational specialty coffee market (Boudreaux, 2011). At the time, 90% of the Rwandan cof-fee was dry-processed and thus classified as low-quality (MINAGRI and MINICOM, 2008). Thegovernment also liberalized the coffee industry and removed barriers to trade (Boudreaux, 2011).

8The dried beans are called parchment coffee.9Farmers clean cherries using rocks and then dry the beans on mats (Macchiavello and Morjaria, 2020).

10In 2012, wet-processed (also called as fully washed) coffee export gate prices are roughly 40% higher than fordry-processed coffee in Rwanda (Macchiavello and Morjaria, 2020).

8

The reforms in the coffee industry changed the incentives of individuals in the coffee valuechain. First and foremost, farmers started to have the incentive to use wet-processing rather thandry-processing since it yields a higher income (Boudreaux, 2011). Exporters compete to sell thewet-processed coffee to foreign buyers. Foreign buyers are interested in buying the Rwandancoffee since the quality increased as the coffee is now wet-processed.

Expansion of the Mills. Between 2000-2006, governmental institutions collaborated with USAID,universities in Rwanda and U.S., and private sector partners under the Partnership to EnhanceAgriculture in Rwanda through Linkages (PEARL) project. The project helped farmers to establishcooperatives, find loans and build mills in their communities in few locations. After 2006, farmerscontinue to establish cooperatives and build mills in their communities across the country. Thenumber of mills expanded rapidly between 2005-2011. Figure 1 shows the expansion of the millsin Rwanda over the last decades. In 2005, the total number of mills was 49. In 2011, the numberquadrupled to 197.

Figure 1: Mill Expansion in Rwanda

Mill Suitability. There is geographical variation in mill locations. Figure 2 visualizes the spatialvariation. The expansion of the mills are mostly concentrated in the areas that had a high numberof coffee trees in 1999, before the adoption of National Coffee Strategy in 2002.

The number one condition for mill suitability is the area having a sufficient amount of coffeetrees (Schilling and McConnell, 2004). This is mainly because the harvested coffee cherries should

9

Figu

re2:

Exp

ansi

onof

Mill

sin

Rw

anda

Not

e:T

hem

aps

are

cons

truc

ted

byco

mbi

ning

data

onm

ills,

The

Rw

anda

nC

offe

eC

ensu

s19

99,2

003,

2009

and

2015

,The

Rw

anda

nG

eoPo

rtal

spat

iald

ata

onse

ctor

boun

drie

s.

10

be transported to the mills within two hours of harvest. Otherwise they will rot (Schilling andMcConnell 2004, Macchiavello and Morjaria 2020).11 An overwhelming majority of the countryis suitable for coffee cultivation.12 Figure A.1 visualizes FAO-GAEZ coffee suitability index andconfirms that the country is suitable for coffee production. Arabica coffee, the dominant plant typein Rwanda, grows best at an altitude over 1,000 meters (3,000 and 6,000 feet) and at an averagetemperature between 15 and 24 °C (59 and 75 °F). Nicknamed as “The Land of the ThousandHills”, Rwanda’s hillsides provide the ideal conditions to cultivate coffee.

Figure 2 and Figure A.1 show that over time, the number of coffee trees increased also inthe coffee suitable areas that did not have coffee trees in 1999. However, specifically the rapidexpansion between 2005 and 2010 is concentrated in the areas where there were already coffeetrees back in 1999. This is mainly because it takes at least 3-5 years for a newly planted coffeetree to produce coffee beans. Thus, a newly planted coffee tree that is planted the day the NationalCoffee Strategy is adopted will start to produce coffee cherries between 2005-2007.

In Section 5.1, I investigate the determinants of a mill opening and present supporting evidencethat conditional on agricultural conditions related to mill suitability, the timing and location of amill opening is random.

3.3 Women’s paid employment in the mills

The division of labor in the coffee value chain is gendered. This holds true not just for Rwanda butin coffee producing countries in the world in general. According to International Trade Forum’s(ITC) survey that was conducted in 15 coffee producing ITC member countries -most of them aredeveloping countries and Rwanda is one of them- women mostly take place in the earlier laborintensive steps in the value chain compared to the latter steps. Survey results is reported in Table1. Women constitute 70% of the workforce in harvesting and processing. Only 10% of themare doing sale activities like in-country trading or exporting (Scholer, 2008). Drying and sortingare female-dominated tasks and marketing and selling the product are male-dominated. As coffeetransforms from a commodity into a value-added product tasks in the coffee value chain becomemale-dominated (SCAA Sustainability Council, 2015).

Before the expansion, the gendered division of labor generated a disparity in generating laborincome in Rwanda. Although women worked in labor intensive stages, men receive and control theincome. Women do not have their personal income and they depend on their husbands financially.

11Most farmers carry their product to mills by walking.12Coffee plants (coffea) is found in the tropical areas of Africa, South America and Asia, which is termed as “the

coffee/bean belt”. The belt provides the necessary conditions for the coffee plants to grow. Such conditions includespecific ranges of temperature, rainfall, altitude and soil characteristics.

11

Tasks in the value chain Participation Variation (Min-Max) Typical Participation

Fieldwork 10-90% 70%Harvest 20-80% 70%Sorting 20-95% 75%In-country trading 5-50% 10%Export 0-40% 10%Other (certification, laboratories) 5-35% 20%Source: International Trade Forum, 2008

Table 1: Women’s Participation in the Coffee Value Chain in Coffee Producing Countries

With the expansion, mills demand seasonal wage labor for drying and sorting tasks.13 Womentransition from being unpaid family workers to wage workers in the mills. What is unique aboutthe context is that women are doing the exact same tasks as before the expansion, drying andsorting, but now they are paid for their labor by the mills and generate personal income. Sincemen were already generating income from their labor (selling the product), there is no change intheir type of earnings. However, I will show that they generate a higher income from selling theirproduct. Before the expansion, men sell dry-processed coffee beans to middlemen in the domesticmarket for a very low price. With the expansion, they sell their cherries to mills for a higher price.14

In 2012, authors of Macchiavello and Morjaria (2020) surveyed the coffee farmers who sellcherries to a coffee mill in Rwanda where the survey results confirm that a mill opening affectswomen and men differently. On average, a mill buys cherries from 400 farmers. 71% of the surveyrespondents are men, which confirms that selling cherries is a male dominated task.15 Moreover,55% of the farmers that answered the survey are coffee cooperative members. 50% of the mills areowned/run by coffee farmer cooperatives. If a mill is owned/run by a cooperative, it means that thecoffee farmers in a region gather together, establish a cooperative and build a mill. Based on thenumbers, typically, the cooperative member farmer husband sells coffee cherries to a mill and hiswife works in the mills during the harvest season, March-July.16

13Within the harvest months, in a given day, multiple farmers bring their cherries to a mill. The cherries arecombined together and washed with the machinery in the mills. In the following days, mills demand paid labor for thefemale-dominated drying and sorting tasks where women can work for.

14Macchiavello and Morjaria (2020) surveys the coffee farmers in Rwanda in 2019. The survey asks the relativeprofitability of the two processing methods. 98% of the farmers reported that selling cherries to mills are moreprofitable than home processing.

15Based on my data, the remaining percent is in line with the share of farmer women who are married to non-farmerhusbands or are widows.

16It has also been documented that the farmers (husbands) who sell to the mill may receive a second-paymentfrom the mills after the harvest, around November. This is when export contracts are realized and mills receive apremium/payment from the countries that the processed coffee is exported (Macchiavello and Morjaria 2020, Church2018). Although 80% of the survey respondents expressed that they received a second payment in the past, the amountof the payment is not big. Amounts typically are 5%–10% of total payments during the harvest season (Macchiavello

12

4 Data

4.1 Panel of Mills

I combine multiple data sources to create a panel of mills.

Mills Data. The Rwanda GeoPortal provides data on the universe of mills as of 2014 in Rwanda.The data is geocoded and includes information on the characteristics of the mills like owner, num-ber of drying tables etc. I combine the data with Macchiavello and Morjaria (2020)’s data thatincludes the year of operation for each mill between 1995-2012.17 By combining the datasets, Icreate a panel of mills that has information on GPS coordinates, year of operation and characteris-tics of every mill in Rwanda.

Spatial Data, Coffee Census and FAO-GAEZ Suitability Index. I spatially match the mills datawith the maps of different geographical boundaries to find the sector, district and province a millis located in. Then I match the data with several rounds (1999, 2003, 2009 and 2015) of RwandanCoffee Census. The coffee censuses provide information on the universe of coffee trees in Rwandaat the sector level. I also match the mills panel with FAO’s Global Agro-Ecological Zones (FAO-GAEZ) coffee suitability index for 1980-2010. The index provides a coffee suitability score forRwanda at 9 km2 resolution. I aggregated the index at the sector level. More details on the dataand summary statistics are available in Appendix C.1.

4.2 Individual and Household Level Data

Rwandan Demographic Health Surveys. As a first step of my analysis, I show the impact ofthe expansion of the mills on paid employment using the Rwandan Demographic Health Surveys(DHS). DHS are nationally representative, cross-section individual and household level surveysthat are conducted in developing countries every 5 years. I use 2005, 2010/11 and 2014/15 cyclesfor my analysis. The surveys collect demographic and health information from women aged 15-49and men aged 18-59. The data also provide household member information, which enable me tolink couples to each other for my analysis.

The information on individuals’ employment and type of earnings is collected retrospectivelyduring the individuals’ interview: individuals are asked to answer whether they are employed forthe last 12 months and if so, whether they worked for cash.18 I create a binary variable which takes

and Morjaria, 2020).17The number of mills opened between 2012-2014 is very small. I coded those mills’ year of operation as 2013.18Due to retrospective questions, DHS data enables me to observe four harvest years in total, 2004, 2005, 2010 and

13

the value one if an individual worked for cash in the last 12 months. Nationally, 88% of womenworked in the past 12 months, where only 39% of them worked for cash. I also use other labormarket outcomes including occupation.

Domestic violence categories (physical, sexual and emotional) are classified by DHS with re-spect to World Health Organization (WHO) guidelines. I create a binary variable which takes thevalue one if a partnered woman experienced physical or sexual domestic violence in the last 12months. 34% of women self reported experiencing domestic violence in the past 12 months.

The surveys also collect GPS coordinates for every cluster of households. Using the GPScoordinates, I spatially merge the DHS data with the mills data. Due to the random displacement(to main confidentiality of respondents), rural clusters contain a minimum of 0 and a maximumof 5 kilometers of positional error. Thus, GPS displacement may lead to measurement error andcan bias the results (Perez-Heydrich et al., 2013). In order to reduce distance measurement error,I follow Perez-Heydrich et al. (2013), which suggests using a buffer distance rather than a closestdistance when using a distance measure. All the distances calculated in the paper are based on abuffer distance measure. I revisit this issue in the robustness checks section.

I restrict the sample to women who are partnered (civil marriage and those who are livingtogether) before the expansion of the mills to avoid changes in the marriage market matching.

Integrated Household Living Conditions Surveys. I complement my analysis using IntegratedHousehold Living Conditions Survey (EICV). EICV is nationally representative, cross-section in-dividual and household level survey that is conducted in Rwanda. I use 2005, 2011 and 2013/14cycles for my analysis. The surveys collect demographic and socioeconomic information fromhouseholds. I use last daily labor income amount (cash) to further investigate the impact of themill expansion on the labor market and uncover the mechanisms behind the effects on domesticviolence. On average women earn less daily labor income compared to their husbands.

The surveys collect district of residence as the smallest geographical unit for each individ-ual/household. There are in total 30 districts in Rwanda. I aggregate the panel of mills at thedistrict level and merge it with individual and household level variables for my analysis. I againrestrict the sample to women who are partnered (civil marriage and those who are living together).More details on EICV and DHS data and summary statistics are available in Appendix C.2.

2014.

14

4.3 Administrative Hospital Level Domestic Violence Data

In order to investigate the impact of providing employment opportunities to women on domestic vi-olence, I also use confidential, administrative, geocoded data on the universe of public/district hos-pitals from the Rwandan Ministry of Health (MOH). Rwandan Health Management InformationSystem (HMIS) data is a monthly district hospital level data on hospitalizations between January2012 to December 2020. I focus on the years between 2012-2014 for my analysis.

There are in total 42 district hospitals in Rwanda. Due to the institutional structure of thehospitals, the data is on the universe of official domestic violence reports in Rwanda. It collectsinformation on the number of individuals (both women and men) who show symptoms of physicaland sexual violence for different age groups (10-18, older than 18). Unfortunately the data does notprovide information on the marital status of the patient. In order to create a measure of domesticviolence, I focus on the gender based violence reports of individuals who are older than 18 yearsold. This is because 70% of women over 18 are married in Rwanda.

I construct a binary variable coded as 1 if a hospital had hospitalizations due to physical orsexual violence for women older than 18 in a month and 0 otherwise. This creates a non self-reported measure of domestic violence. 79% of the hospitals have hosted at least a domesticviolence patient (GBV patient older than 18) in a given month. Geocoded nature of the data enableme to match this measure of domestic violence with the panel of mills. More details on the dataand summary statistics are available in Appendix C.3

4.4 Measuring Exposure to Mills: Treatment and Control Groups

Throughout the paper, the treatment group consists of entities (woman/husband/hospital) who areexposed to the mills where the control group consists of the entities who are not. I construct twomeasures of exposure to the mills, being in the catchment area of a mill and mills per capita in adistrict. I also construct multiple control groups with respect to the measures I create.

The first mill exposure measure, being in the catchment area of a mill, is used when the out-come variables are created with geo-coded datasets (DHS and HMIS). The measure uses the GPScoordinates of the mills and DHS clusters/hospitals and calculates whether the clusters/hospitalsare within the catchment area of a mill. In most cases, mills are located within a radius of 3–5km away from the farmers (AgriLogic, 2018). On average, managers of the mills in Rwanda re-port that the catchment areas have a radius of approximately 4.5 km (Macchiavello and Morjaria,2020). I create a buffer with 4 km radius centered around the mills to construct the catchment area(treatment group). I define a DHS cluster/hospital within the catchment area of a mill, if the GPS

15

coordinate of the DHS cluster/hospital falls inside the buffer around a mill.

When using DHS data, I consider 2 potential control groups in my analysis. The first controlgroup consists of the DHS clusters that are outside of the catchment area of the mills (4 km buffer)but are located within the same district with the mills. I define using this control group as the“within district approach”. An alternative control group consists of only the DHS clusters that arewithin the donut area between 4 and 8 km from the mills. Since I restrict the control group onlyto DHS clusters that are within the surface area of a donut, I define using this control group asthe “donut approach”. Figure 3 shows a visual representation of the treatment and the 2 controlgroups using a mill in Nyarugenge district and DHS clusters. In both maps, the circular area (bufferaround the triangle/mill) is the treatment group. In the top map, which visualizes the within districtapproach, everywhere outside the buffer characterizes the control group. In the bottom map, whichvisualizes the donut approach, only the donut area constitutes the control group.

Figure 3: Visualization of the Treatment and Control Groups using Nyarugenge District: WithinDistrict and Donut Approach

For HMIS (hospital) data, the control group consists of the hospitals that are outside of thecatchment area of the mills but are located within the same province with the mills.This choice isdue to the sample size and locations of the hospitals. There are only 42 hospitals in the countryand majority of the hospitals who are not in the catchment area of a mill are located outside of thesame district with the mill as well as the donut area between 4 and 8 km from the mill. Figure 4shows a visual representation of the treated and control hospitals. The ones in the catchment areasare treated hospitals where the ones outside (within the same province) are control hospitals.

The underlying assumption I make in the hospital specification is that households within (out-side of) the catchment area visit a hospital inside (outside) of the catchment area rather than theone outside (inside) due to the hospital’s proximity. This is a plausible assumption since the costof visiting a hospital far away in case of a health problem is more costly compared to visiting the

16

Figure 4: Visualization of the Treated and Control Hospitals

one that is closer (within the catchment area). I create 4 km buffers around the hospitals to checkwhat percent of the buffers around the control hospitals do not intersect with mill buffers. Thelogic here is that if the rate of no intersection is high, then most of the control hospitals are farenough from the households within the catchment areas. Thus, those households who are withinthe catchment areas plausibly go the hospitals within the catchment areas and the households out-side of the catchment areas go to the ones outside due to proximity. I find that 70% of the controlhospital buffers do not intersect with mill buffers. In any case, if a household within the catchmentarea of a mill visit a hospital outside of the catchment area, this suggests that my results constitutea lower bound.

The second mill exposure measure, log of mills per capita, is used when the outcome variablesare created with a dataset that is not geo-coded (earnings data, EICV). The measure is the log oftotal number of mills in a district in a given year divided by the number of working age individualsliving in the district in that year.19

5 Empirical Specification

This section proposes the identification strategy to estimate the causal impact of a mill opening onwomen’s labor market outcomes and domestic violence. I use a difference-in-differences strategythat uses the differential timing of and spatial variation in a mill opening.

19Individuals who are aged 19-59 are selected in EICV to match the age profile in the DHS sample.

17

5.1 Identifying Assumptions and Threats to Identification

The key identifying assumption of the empirical strategy is that the average outcomes for the treat-ment and control groups would have parallel trends in the absence of treatment, a mill placement.A mill opening at a specific location in a given year is assumed to be uncorrelated with other de-terminants of changes in the outcome variables (women’s labor market outcomes and domesticviolence) over time (the treatment is not endogenous). If a mill placement is endogenous, then theparallel trends assumption is violated. This is because in counterfactual, the areas who are exposedto the mill would have diverged anyway, regardless of the mill.

To provide suggestive evidence in favor of my identifying assumption, I estimate, at the sec-tor level, the determinants of having the first mill during the period when mills expanded rapidly,between 2005-2011, and having a mill by the end of my sample, 2014. Firstly, I focus on thevariables that are related to coffee cultivation including the historical number of coffee trees andFAO-GAEZ coffee suitability index. I also include the shares of women who are unpaid familyworkers, self-employed and completed primary school at the sector level before the expansion ofthe mills as well as the share of men who completed primary school in the regression. This is toconfirm that mill placement is uncorrelated with factors that can affect the evolution of female paidemployment and domestic violence rates in Rwanda. I use the Rwanda Population and HousingCensus 2002 to create the variables. I also included total population and female population at thesector level. Unfortunately, I am unable to include domestic violence rates before the expansionin the regression, since domestic violence data started to be collected in 2005, after the expansionstarted. However, since education (women and men’s) are correlated with the probability of ex-periencing domestic violence, including them in the regression constitutes a test on whether theplacement of the mills is correlated with the evolution of domestic violence rates in Rwanda.

I also include other variables related to promoting female empowerment including share ofwomen in a consensual union, polygamous marriage, without assets and number of daughtersper woman. If these variables are uncorrelated with mill placement, then, it provides supportingevidence that the mills are not placed to promote female empowerment and hence mill placementis not endogenous. Moreover, I include genocide intensity since it is also a variable that is foundto impact the probability of experiencing domestic violence in Rwanda (La Mattina, 2017). It isalso possible that the government or NGOs are more likely to financially support the opening ofa mill in genocide intense areas to promote female empowerment that can make the opening of amill endogenous. District fixed effects are included in all specifications.

I report the results in Table 2. The results provide support for the evidence that variablesrelated to coffee, the number of coffee trees in 1999 and FAO-GAEZ coffee suitability predicts

18

mills placement. Specifically, sectors that have a higher number of coffee trees in 1999 are morelikely to have their first mill between 2005-2010. Those areas are also more likely to have a millby 2014. Similarly, sectors with a higher FAO-GAEZ coffee suitability index are more likely tohave their first mill between 2005-2010. Based on results, historical number of coffee trees aremore likely to be correlated with the rapid expansion of the mill compared to FAO-GAEZ coffeesuitability index. This is not surprising. Before the expansion, not every coffee suitable areain Rwanda had coffee trees and it takes minimum 3-5 years for a coffee tree to produce coffeecherries. Thus, it is likely to observe the rapid expansion of the mills between 2005-2010 in theareas that already have coffee trees before the expansion, like 1999.

All variables other than the number of coffee trees in 1999 and FAO-GAEZ coffee suitabilityindex are statistically insignificant. Specifically, pre-expansion female unpaid employment andfemale, male primary education rates do not predict mill placement. None of the variables relatedto female empowerment is statistically significant. Genocide-intensity does not have a statisticallysignificant impact on mill expansion. I also test the joint significance of the relationship betweena mill opening and non-coffee variables. For both regressions, I cannot reject the null hypothesis(p-value=0.4954 and p-value=0.6026). These results suggest that mills are not placed to promotefemale empowerment. They are uncorrelated to the factors that can have an impact on the evolutionof female paid employment and domestic violence over time. Thus, conditional on mill suitabilitybased on coffee related variables, opening of a mill is assumed to be random. I control for historicalnumber of coffee trees and FAO-GAEZ coffee suitability index in all of my specifications in thepaper.

In Section 9, I also perform an event-study analysis for all of my main outcome variables thatexploits the variation in the number of years couples are exposed to a mill. There is no pre-trendsin outcomes.

5.2 Paid Work and Self-Reported Domestic Violence

In order to investigate the impact of a mill opening on women’s and their husbands’ outcomes, Iestimate the specification below:

Yist = β0 +β1Millist +Xistφ +λc +ωm +αs + γdt +(Xs× t)θ + εist . (1)

The dependent variable Yist is the outcome of interest of woman i (or the husband of woman i), insector s and at year t. Millsist is a binary variable coded as 1 if woman/husband i at year t resideswithin the catchment area of a mill and zero otherwise. I have a rich set of individual controls,

19

Table 2: Sector Level Baseline Characteristics that Predict Mill Opening

(1) (2)First Mill in 2005-2010 Mill by 2014

Log Coffee Trees in 1999 0.03*** 0.04***(0.01) (0.01)

FAO-GAEZ Coffee Suitability Index 0.07 0.03(0.04) (0.04)

Log Population in 2002 -0.95 -0.67(0.74) (0.75)

Log Female Population in 2002 0.91 0.73(0.73) (0.74)

Share of Self-Employed Women in 2002 0.43 0.65(0.55) (0.55)

Share of Unpaid Worker Women in 2002 0.36 0.87(0.58) (0.59)

Share of Primary-Educated Women in 2002 1.75 1.32(1.12) (1.14)

Share of Primary-Educated Men in 2002 -1.00 -1.10(1.21) (1.22)

Number of daughters per Woman in 2002 -0.18 -0.13(0.22) (0.22)

Share of Women in a Consensual Union in 2002 -1.06 -0.57(0.77) (0.78)

Share of Women in a Polygamous Marriage in 2002 -0.02 -1.08(1.94) (1.96)

Share of Women without Assets in 2002 0.41 -0.32(0.60) (0.60)

Age at Genocide among Women in 2002 -0.08 -0.06(0.05) (0.05)

Genocide Intensity Index at the Commune Level 0.00 -0.03(0.04) (0.04)

District FE X X

Number of Observations 348 348Dependent variable mean 0.28 0.33Adjusted R2 0.27 0.33

Note: FAO-GAEZ coffee suitability and genocide intensity index are both standardized. The data is at thesector level. FAO-GAEZ index is originally defined at the 9 km2 resolution. I aggregated the index at thesector level (on average 50 km2 area) for analysis. *** p<.01, ** p<.05, * p<.1

20

Xist, that includes information on women’s/husbands’ occupation, education, religion, numberof children, marital status, age and duration of marriage, residence (rural/urban) and householdwealth. I also control for partner characteristics like partner’s age, occupation and education. λc isthe cohort fixed effects and controls for factors that vary across cohorts. As an example, comparedto younger cohorts, older cohorts grew up during a period when Rwanda did not have pro-womenlaws which may affect their labor market participation and acceptance of domestic violence. ωm isthe year of marriage fixed effects. It controls for time-variant shocks to the marriage market suchas the Rwandan Genocide that is documented to affect marriage quality in Rwanda (La Mattina2017, Sanin 2021).20 αs is the sector fixed effects and controls for time-invariant local observableand unobservable characteristics, such as social norms related to women’s employment, domesticviolence and gender roles.

γdt is the district-by-year fixed effects. This is to control for factors that change over time andacross districts and may determine both a mill opening and female empowerment such as femalepolitical participation. As of 2008, Rwanda is the first country in the world with a female major-ity in parliament. The share of women in local government varies across districts and increasesover time. Districts (akarere) are the geographical units with the highest tier in local government.District councils decide on local development programs within districts. As a possible concern,after an increase in the female political representation in a district council, the local governmentmay financially support cooperatives who want to open mills to promote female paid employment.District-by-year fixed effects control for such scenario. Xs is a vector of baseline geographicalvariables at the sector level such as the historical number of coffee trees in 1991 and FAO-GAEZcoffee suitability index. I interact these initial conditions with linear time trends to allow theirimpact vary over time. The interaction mitigates potential omitted variable bias. This is becausecoffee tree presence in 1999 is correlated with mill openings after 2002 and it may also effect theevolution of female paid employment over time. I cluster standard errors at the sector level.

The main dependent variables for this specification are being employed, working for cash andexperiencing domestic violence in the past 12 months. All outcomes are indicator variables. Asan example, working for cash in the past 12 months variable takes the value one if the respon-dent worked for cash in the past 12 months and 0 otherwise. Domestic violence in the past 12months variable is a combination of physical and sexual domestic violence. It takes the valueone if a partnered woman experienced physical or sexual domestic violence in the past 12 months

20La Mattina (2017) shows that the timing of marriage (before or after the genocide) has an impact on probabilityof experiencing domestic violence in Rwanda. Sanin (2021) investigates the impact of the adoption of the domesticviolence legislation in Rwanda in 2008, which allows women to unilaterally divorce their husbands if their husbandsare violent towards them. After the law, among the women who married after the genocide, the divorce rates increasemore and sexual domestic violence rates increase less in the formerly genocide-intense areas, where women are morelikely to be in violent marriages. No effect is observed for women who married right before the genocide.

21

and 0 otherwise. I also use dependent variables like occupations (manager, sales, agriculturalself-employed, agricultural employee, skilled manual, unskilled manual) and household decision-making (whether large household purchases are decided by women alone or jointly with husbandcompared to husband alone).

The coefficient of interest is β1, which identifies the impact of a mill opening on the outcomevariables. The treated group consists of woman/husband who reside in the catchment area of amill. The control group either consists of woman/husband who do not reside in the catchment areaof a mill within the same district (within district approach) or woman/husband who reside withinthe donut area between 4 and 8 km from a mill (donut approach). Tables A.11 and A.12 examinewhether predetermined covariates including women’s agricultural occupation, education, religion,marriage type, age at first marriage, partner characteristics and living conditions (electricity, ce-ment floor in the household) are balanced across treatment and control groups. I perform the testby estimating the specification given in equation 1 using each of the predetermined covariates asthe dependent variable. None of the estimates are statistically significant which suggests that thebaseline characteristics are balanced across the treatment and each control group.

I also perform a placebo test to show that outcome variables are balanced across treatmentand control groups before a mill opening. For this test, I drop the women who reside within thecatchment area of a mill in a given year (2004, 2005, 2010 and 2014) from the sample. I constructthe treatment group as women who live in the areas that do not have a mill yet. The areas willreceive a mill and become catchment areas in the upcoming years. Thus, I am falsely assuming

that the women who reside in those areas are exposed to a mill and in the treatment group. Thecontrol group consists of women who live in the areas outside of the future catchment areas. Iagain use both the within district and donut approaches. Women in both groups self-report theirlabor market outcomes and domestic violence experience for a given year before a mill opening.Since both groups are not exposed to a mill, I expect to see no statistical differences betweenthem. I perform the placebo test by estimating the specification given in equation 1 using theaforementioned treatment and control groups. None of the estimates are statistically significantwhich suggests that the outcome variables are balanced across treatment and control groups beforea mill opening.

5.3 Earnings

I also test the impact of a mill on earnings (women and their partners’ each).21 Unfortunatelythe data on these variables is not geo-coded and the smallest geographical unit in the data is the

21I use earnings and labor income interchangeably.

22

Table 3: Placebo Test: Outcome Variables Before a Mill Opening

Within District Donut

(1) (2) (3) (4) (5) (6)Work Cash Violence Work Cash Violence

Mill 0.01 -0.02 -0.00 0.00 -0.03 0.01(0.01) (0.04) (0.07) (0.01) (0.04) (0.07)

Observations 6472 6475 2717 3105 3106 1348Dependent variable mean 0.99 0.36 0.35 0.98 0.36 0.34

Note: Robust standard errors clustered at the sector level are in parentheses. 4 km catchment area is usedfor the treatment group. The estimates are based on DHS data and estimated with the main specificationpresented in Section 5.2.1 *** p<.01, ** p<.05, * p<.1

district. Thus, in order to estimate the impact of the mills on those variables, I employ the empiricalspecification below:

Yidt = β0 +β1Millidt +Xidtφ +λc + γdt +(Xd× t)θ + εidt . (2)

The dependent variable Yidt , is the logarithm of last daily labor income of partnered woman/husbandi, in district d and at year t for estimating the impact on earnings. Millsidt is the logarithm of thetotal number of mills per capita in the district of residence of a woman/husband i at year t. Xidt

is the same set of controls in the main specification. λc is the cohort fixed effects.22 γdt is thedistrict-by-year fixed effects. Xd is a vector of baseline geographical variables at the district levelincluding the historical number of coffee trees in 1999 and FAO-GAEZ coffee suitability index.I interact these initial conditions with linear time trends, to allow their impact vary over time. Icluster standard errors at the district level.

The coefficient of interest is β1, which identifies the impact of an increase in log mill per capitaon the outcome variables.

5.4 Monthly Hospitalizations due to Domestic Violence

A unique feature of this paper is that I investigate the relationship between paid work opportunitiesand domestic violence using both annual self-reported and monthly administrative data. Using theuniverse of monthly hospitalizations due to domestic violence in Rwanda, I test whether a mill

22I am unable to control for year of marriage fixed effects since the data does not have such information. However,I am controlling for cohort fixed effects.

23

affects domestic violence when it is in operation. This variation is due to the fact that a mill isoperating only during the harvest period: March-July. I use a specification at the hospital level. Toestimate the impact of a mill on hospitalizations due to domestic violence, I estimate the empiricalspecification below:

Yhdtm = β0 +12

∑m=1

Millhd×βm1[τ = m]+Xhtφ +λh +αd +σm + γpt +(Xd× t)θ + εhdtm. (3)

The dependent variable Yhdtm is the monthly hospitalization outcome due to domestic violencein hospital h, in district d, in year t and at event-time m. Millshd is a binary variable codedas 1, if hospital h is within the catchment area of a mill and zero otherwise during the sampleperiod, 2012-2014. Being in the catchment area of a mill did not change between 2012-2014 forthe hospitals. Thus, Millhd is a time-invariant characteristic. It is interacted with event-monthdummies, 1[τ = m], to investigate the dynamic impact of a mill during the harvest period (March-July), the months mills operate. τ denotes the event-month. τ = 3, March, represents the monthharvest period begins and mills start to operate. For 3 ≤ m ≤ 7, March-July, τ = m representsthe months mills are operating. For m < 3, τ = m represents the months before a mill’s month ofoperation. The omitted category is τ = −1, February, which means that the dynamic impact ofbeing exposed to a mill is estimated with respect to one month prior to a mill’s month of operation.

Xist is the set of hospital level time-varying controls related to gender based violence specifichospital quality. It includes information on whether gender based violence patients are referred forcare to a health facility with higher level of resources. It is a proxy for poor hospital quality. Xist

also includes information on whether gender based violence victims are referred to the hospital bythe police and community health workers. They are proxies for high hospital quality.23 λh is thehospital fixed effects which controls for any hospital specific characteristic that is fixed over timelike its location. αd is the district fixed effects. District is chosen for the level of geographical unitsince the unit of observation is a district hospital. σm is the month fixed effects and controls formonth specific trends. γdt is the province-by-year fixed effects. I allow year fixed effects to differby province, one unit higher than the district.24 This way, I am comparing the hospitals who arein the catchment area of a mill to the ones that are not, within the same province. Hospitals whoare not within the catchment area of a mill within the same province constitutes a more accuratecontrol group. Xd is the vector of baseline geographical variables at the district level interacted

23Being a hospital that is referred to a gender based violence victim by the police and community health workersmay show that the hospital is known for the resources and knowledge related to gender based violence cases. Moreover,referrals may also show that the hospital is in collaboration with the police and community health workers.

24Unfortunately there is not enough observations/hospitals to have district-by-year fixed effects. Adding district-by-year fixed effects mean comparing the hospitals who are in the catchment area of a mill to the ones that are notwithin the same district. Most of the districts have only one district hospital.

24

with linear time trends. I cluster standard errors at the district level.

The main dependent variables for this specification are whether the hospital had hospitaliza-tions and deaths due to gender based violence for women/men older than 18 and aged 10-18.Variables that focus on individuals older than 18 capture domestic violence. This is because ac-cording to DHS and census data, majority of the individuals who are older than 18 are married inRwanda. Variables that focus on individuals aged 10-18 capture gender based violence not due toa partner. Moreover, that age group do not work in the mills. Such variables are used as placebooutcomes as a robustness check.

All (violence) outcomes are indicator variables. As an example, monthly hospitalizations dueto gender based violence for women older than 18 variable takes the value 1 if a hospital had ahospitalization due to gender based violence for a female victim older than 18 in a given monthand 0 otherwise. Dummy variables rather than the logarithm of hospitalizations are used since thenumber of hospitalizations for gender based violence is low in a month (close to 1). Only when Iam performing the placebo test, when the dependent variable is hospitalizations for bone and jointdisorders other than fractures, I use the logarithm of hospitalizations. This is due to the fact thatmean hospitalizations for such diseases is way bigger than one, 46.

The coefficients of interest are βm’s when 3 ≤ m ≤ 7 (March-July). For 3 ≤ m ≤ 7 and agedolder than 18 category, each βm provides the change in hospitalizations due to domestic violencewithin the cathment area of a mill relative to the hospitalizations that are not, within the sameprovince, during a mill’s month of operation, relative to one month before the operation. Theharvest season is at peak during 5 ≤ m ≤ 7, May-July, where majority of the farmers work in themills. Thus, I specifically expect to see statistically significant estimates during those months.Since m = 1 (January) and 8≤ m≤ 12 (August-December) respectively represent the months be-fore and after the harvest season when the mills do not operate, I expect to see small and statisticallyinsignificant estimates for those months.

6 Results

6.1 Employment and Self-Reported Domestic Violence

I first estimate the impact of mill exposure on women’s employment, type of earnings and self-reported domestic violence using equation 1. I show that a mill opening increases women’s paidemployment and decreases women’s likelihood of experiencing domestic violence in the past 12months.

25

Table 4 presents the results of estimating the impact of exposure to a mill on women’s proba-bility of working in the past 12 months, her type of earnings being cash in the past 12 months andself-reporting domestic violence in the past 12 months using different control groups. To measuremill exposure 4 km catchment area is used. The first three columns represents estimating equation1 when the control group is defined as DHS clusters that are outside of a mill’s catchment areabut within the same district the mill is located (within district approach). Columns 4-6 representsestimating equation 1 when the control group consists of DHS clusters that are within the donutarea between 4 and 8 km from a mill (donut approach). All estimations include individual con-trols, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-year fixedeffects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills.

When within district approach is used for the control group, upon a mill opening, being exposedto a mill increases the probability of working for cash in the past 12 months by 7 percentage points(p-value= 0.001) where the probability of working in the past 12 months remain unchanged. Theestimated impact represents an increase of 18% with respect to the sample mean (0.40). Moreover,being exposed to a mill decreases the probability of experiencing domestic violence in the past 12months by 9 percentage points (p-value= 0.01). The estimated impact represents a decrease of26% with respect to the sample mean (0.35).

When using the donut approach, upon a mill opening, being exposed to a mill increases theprobability of working for cash in the past 12 months again by 7 percentage points (p-value=0.003) where the probability of working in the past 12 months remain unchanged. The estimatedimpact represents an increase of 16% with respect to the sample mean (0.45). Moreover, beingexposed to a mill decreases the probability of experiencing domestic violence in the past 12 monthsby 8 percentage points (p-value= 0.06). The estimated impact represents a decrease of 22% withrespect to the sample mean (0.37). In summary, results continue to be statistically significant andvery similar even when the control group is restricted only to the clusters within the donut areabetween 4 and 8 km from a mill. There is a statistically significant increase in the probability ofworking for cash and a decline in the probability of experiencing domestic violence for women.25

Table 5 presents the results of estimating the impact of exposure to a mill on women’s hus-bands’ probability of working in the past 12 months and their type of earnings being cash in thepast 12 months for both control group approaches. There is no statistically significant change inthose variables. I also estimated the impact of mill exposure on women’s and their husbands’ occu-

25I also check alcohol consumption among husbands. DHS 2005 and 2014 ask for such information. 61% ofhusbands in my sample (based on DHS 2005, 2014) gets drunk sometimes and 20% gets drunk very often (rest isnever drunk). The means are similar across within and outside of the catchment area. There is no change in drinkingbehavior based on DID using 2005 and 2014.

26

Table 4: Effect of Mill Exposure on Women’s Employment, Type of Earnings and Self-ReportedDomestic Violence in the Past 12 Months


(1) (2) (3) (4) (5) (6)Work Cash Violence Work Cash Violence

Mill -0.00 0.07*** -0.09*** -0.00 0.07*** -0.08**(0.01) (0.02) (0.03) (0.01) (0.02) (0.04)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All dependent variables are measured forthe past 12 months. The catchment areas is constructed by buffers around the mills with a 4 km radius. ***p<.01, ** p<.05, * p<.1

Table 5: Effect of Mill Exposure on Husband’s Employment and Type of Earnings in the Past 12Months


(1) (2) (3) (4)Work Cash Work Cash

Mill -0.00 0.03 -0.01 0.02(0.02) (0.02) (0.02) (0.02)

Observations 4099 3548 2168 1961Dependent variable mean 0.87 0.81 0.90 0.82

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists of thehusbands of the women in the sample. All dependent variables are measured for the past 12 months. Thecatchment areas is constructed by buffers around the mills with a 4 km radius. *** p<.01, ** p<.05, *p<.1

27

pation. It is possible that a mill opening can induce individuals to sort into the agricultural sector.Results are reported in Tables A.13, A.14, A.15 and A.16. The first two tables use the withindistrict approach and the last two use the donut approach. Each column represents one occupationwhere the occupations are managers, sales, agricultural self-employed, agricultural employee andskilled and unskilled manual. For women, exposure to a mill do not induce a change in occupa-tions. The dominant occupation is the agricultural self-employed category where 80% of womenhave such occupation. Being a farmer constitutes an example of working as self-employed in agri-culture. For husbands, there is a statistically significant increase in the sales category. However,there is not a statistically significant impact of mill exposure on husbands being self-employed inthe agricultural sector. Being self-employed in the agricultural sector is also the dominant occupa-tion for husbands where 66% of the husbands have such occupation. Overall, individuals who areexposed to a mill do not sort into being self-employed in agriculture (become farmers).

According to the results, women and their husbands were already self-employed in the agricul-tural sector (farmers). They were already working in the past 12 months. Being exposed to a mill,did not induce a shift on these variables. However, exposure to a mill changed the type of earningsof women. Women start to be paid cash for their labor in agriculture. Their husbands were alreadyworking for cash and being exposed to a mill did not change their type of earnings.

6.2 Earnings

I also estimate the impact of a mill opening on women’s and their husbands’ earnings using equa-tion 2. Using earnings data, I show that a mill opening increases both women’s and their husbands’last daily earnings.

Table 6 presents the results of estimating the impact of a mill opening on women’s and theirhusbands’ log of last daily earnings.26 Since EICV data is not geocoded, mill exposure is definedas the logarithm of mills per capita in a district. Columns 1-2 present the results for the wholesample. Columns 3-4 present the results among couples working in agriculture. Among all sample,1% increase in mills per capita increases women’s and their husbands’ last daily earnings by 1.8%and 2.3% respectively. Among couples working in agriculture, an increase in mills per capita alsoincreases the log of last daily earnings for each spouse. These results show that being exposed to amill upon a mill opening increases women’s and their husbands’ earnings.

It should be noted that the results are based on data where I observe the labor income of womenwho work for pay (positive, not missing labor income). The missing labor income observations

26Majority of the earnings is received on a daily basis in this context due to the dominant occupation is beingagricultural. Thus, I use last daily earnings as my dependent variable.

28

Table 6: Effect of Mill Exposure on Log of Last Daily Earnings

All Sample Occupation: Agriculture

(1) (2) (3) (4)Woman: Log of

Last DailyEarnings

Husband: Log ofLast DailyEarnings

Women: Log ofLast DailyEarnings


Log of Mills percapita in the District 1.80*** 2.37*** 2.10*** 2.62***

(0.18) (0.20) (0.13) (0.14)


Note: Robust standard errors clustered at the district level are in parentheses. All estimations includeindividual controls, cohort fixed effects, district-by-year fixed effects, linear time trends interacted withbaseline district level characteristics. Sample consists of partnered women (and their husbands) whomarried before the expansion of the mills. Since EICV is not geocoded, the mill variable is log of mills percapita in a district in a given year. In Columns 3-4, the sample consists of women and their husbands whoreported their occupation as agricultural. *** p<.01, ** p<.05, * p<.1

belong to women who do not work or do not work for pay. Their labor income is zero eitherway. As a first robustness check for my results on earnings, I code the missing values as zero andtransform the dependent variables by using an inverse hyberbolic sine (IHS) function. Unlike in alog transformation, zeroes are defined in an IHS transformation. Table A.17 presents the results.Estimates are in line with the main results.

Since the data on earnings are available only for a self-selected group of women (women whowork for pay), sample selection bias is a potential concern. In order to address this concern, asa second robustness check, I use the two-step Heckman estimator. An exclusion restriction, avariable that affects women’s decision to work for pay (her reservation labor income), but notthe labor income (labor income offer), identifies the two-step model. I use the number of youngchildren (age 0-5) in the household as such variable in my model. Having young children at homemakes it more costly for women to work due to taking care of children. However, it does not affectthe amount of income received for per unit labor. The variable is commonly used in the literaturethat uses the correction in models on female wages and labor force participation (Mulligan andRubinstein 2008, Blau and Kahn 2017). Table A.18 presents the results. The estimates are stillpositive and statistically significant.

29

6.3 Monthly Hospitalizations for Domestic Violence

I complement my analysis using monthly hospitalizations for domestic violence. Mills providejob opportunities to women only during the harvest months. I show that there is a decline inhospitalizations during the harvest months relative to one month before the harvest season. Yet,the decline is short-lived and not observed during the post-harvest months. The exercise has twoadvantages. First, the monthly results confirm the previous estimates. Moreover, this confirmationis done via administrative records which alleviates concerns regarding the reporting bias in self-reported data. Second, seasonality of the decline in hospitalizations speaks to the mechanisms. Itshows that in a context where wages and household resources are low, access to credit marketsand saving technologies are limited, the effects of providing job opportunities to women may beconcentrated when women have the jobs. If women have better outside options and husbands’non-monetary benefit from violence is smaller due to women’s contribution to household earningsspecifically when women work, it is plausible that the decline in hospitalizations are concentratedduring the harvest months only.

I estimate the dynamic impact of mill exposure on monthly hospitalizations for gender basedviolence using equation 3. Figure 5 plots the coefficient of the interaction terms for every month ina calendar year. Mills operate only during the harvest season, March-July. In a given year, Januaryand February are the pre-harvest period, May-July is the peak of the harvest where majority of theneighboring community around the mills work in the mills. August-December is the post-harvestperiod. In the specification, the omitted category is February, which means that the dynamic im-pact of being exposed to a mill is estimated with respect to one month prior to a mill’s month ofoperation. The figure is for women aged older than 18. Since overwhelming majority of womenaged older than 18 is married in Rwanda, I define hospitalizations for gender based violence forwomen age older than 18 as domestic violence. I find that it is 18 and 14 percentage points lesslikely for a hospital in the catchment area of a mill to have a domestic violence patient in June andJuly respectively (peak of the harvest) compared to one month before the mills’ month of operation(p-value= 0.01 and p-value= 0.06). The estimated impact represents a reduction of 23% and 17%respectively with respect to the sample mean (0.79). There is not a statistically significant changein hospitalizations when mills do no operate (January, August-December). Figure A.4 visualizesthe raw means for hospitalizations for domestic violence. The trend in means are also in line withthe estimates.

Before moving on, one important thing to recall is that the hospitalizations results are basedon the period after the rapid expansion of the mills, during 2012, 2013 and 2014. In those years,there is no change in hospitalizations in the post harvest months relative to February, one monthbefore the harvest season. No change in hospitalizations during the post-harvest months do not

30

mean that the hospitalizations in the catchment areas and non-catchment areas are the same. Asseen in Figure A.4, during post-harvest months, the hospitalization rates in the catchment areas arestill lower compared to the rates outside of the catchment areas.

Figure 5: Dynamic Impact of a Mill Opening on Hospitalizations for Gender Based Violence(Women)

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includehospital controls, hospital fixed effects, district fixed effects, province-by-year fixed effects, linear timetrends interacted with baseline district level characteristics. 4 km catchment area is used. *** p<.01, **p<.05, * p<.1

Figure A.3 shows that there is no changes in hospitalizations for gender based violence amonggirls aged 10-18. I do the same analysis for men and do not find any statistically significant results.I also do not find any statistically significant change in deaths due to gender based violence forboth age groups and genders. I plot the results in Figures A.5 and A.8 respectively. All results arealso shown in Tables A.20.

Placebo Test. As a placebo test, I perform the same analysis for universe of monthly hospitaliza-tions for other diseases, bone and joint disorders other than fractures.27 Examples of such disordersinclude osteoarthritis, gout, rheumatoid arthritis, lupus and bursitis. I report the results in Figure 6

27This means that broken bones, injuries are not included. Thus the these hospitalizations are not related to domes-tic violence.

31

and Table A.21. I find no changes in hospitalizations for these disorders within a year, specificallyduring the harvest months relative to one month before the harvest season.

Figure 6: Dynamic Impact of a Mill Opening on Log Hospitalizations for Bone and Joint Disordersother than Fractures (Women)

Note: Dependent variable is in logs. Fractures (broken bones) are not included. Examples of bone andjoint disorders include osteoarthritis, gout, rheumatoid arthritis, lupus, bursitis. Robust standard errorsclustered at the sector level are in parentheses. All estimations include hospital controls, hospital fixedeffects, district fixed effects, province-by-year fixed effects, linear time trends interacted with baselinedistrict level characteristics. 4 km catchment area is used. *** p<.01, ** p<.05, * p<.1

This rules out the concern that women may go to hospitals less during the harvest season dueto the increased opportunity cost of a hospital visit. For women who are exposed to a mill, a day inthe hospital is costly during the harvest season since it means loss of daily labor income. This canmake women less likely to go to a hospital and drive the decline in hospitalizations for domesticviolence during the harvest months. No change in non-domestic violence hospitalizations suggeststhat the decline in hospitalizations due to domestic violence is not driven by a decreased likelihoodof going to a hospital during the harvest season.

One might argue that the group of women who have bone and joint disorders may be veryold, less likely to work in the mills and thus their opportunity cost of time do not increase withthe harvest season (let’s say it remains the same within a year). If that is the case, no change inhospitalizations for these women is not ruling out the possible scenario for the decline in violence

32

hospitalizations where women go to hospitals less due to the increased opportunity cost of goingto a hospital. I build on medical research to tackle this argument. Medical research suggests thatalthough osteoarthritis (OA), the most common disease in the data, becomes more prevalent asone gets older, age is not a precondition to have the disease. It is also prevalent among occu-pations including farmer, construction worker, miner, that are physically demanding and requireindividuals to use the same bones and joints regularly (Yucesoy et al., 2015). This suggests that thewomen who visit a hospital for bone and joint disorders and domestic violence are plausibly bothworking farmers who have similar opportunity costs of going to a hospital.28 One may also arguethat among farmers in the same age group, women with bone and joint disorders can have a loweropportunity cost of going to a hospital compared to women who experience domestic violence.This low opportunity cost may not increase with the harvest season as well. This is plausibly be-cause those women have a long term disease that decreases their labor productivity regularly andmay make them unable to work efficiently within a year including the harvest months. However,it should be noted that according to data, domestic violence is also a health problem that occursregularly and thus may affect women’s labor productivity consistently. Thus, those two groups ofwomen are comparable and the placebo test plausibly rules out the change in opportunity cost ofgoing to a hospital as a mechanism behind the decline in violence hospitalizations.

7 Conceptual Framework

My conceptual framework guides the mechanisms behind the results and shows how changes inwomen and their husband’s earnings affect domestic violence.

7.1 Setup

Preferences. The household consists of a wife and husband, j ∈ {w,h}. I assume that the prefer-ences can be represented by the utility functions

Uh = yIh +(1− x+dτ)Iw +dα(I)+θh and Uw = (1− y)Ih +(x−dτ)Iw +θw−dvw (4)

where I j for j ∈ {w,h} indicates personal earnings for the wife and husband. I is total householdresources where I = Iw+Ih. Ih > Iw, the husband earns more than the wife as in my data. Accordingto Rwandan social norms, the husband controls household income and household decision-making

28The number of hospitalizations for bone and joint disorders is much higher than the number of domestic violencehospitalizations. It is also plausible that some women visit a hospital for both health problems.

33

(ICRW 2011, Bayisenge 2010, Ya-Bititi et al. 2019). The husband keeps yIh of his personal in-come, y ∈ (0,1), and gives (1− y)Ih of it to the wife for her expenses. The wife keeps xIw of herpersonal income, x ∈ (0,1), and turns (1− x)Ih of it to the husband. In my data, 65% of womendecide how to spend their earnings jointly with their husband, not by themselves, which providessupport for the model. x and y are predetermined in the beginning of the marriage. The maindifference between the husband and the wife is that the husband can extract more money thanhis spouse turns over to him (based on the predetermined x value) by inflicting violence. Mainly,the husband can use violence to get a bigger share of her earnings and maximize his utility frommarriage. Violence decision is parametrized by d ∈ {0,1}. τ ∈ (0,x] is the extraction rate. If thehusband chooses to inflict violence on his wife, d = 1. If he chooses not to be violent, d = 0. τ

captures instrumental violence and shows that violence is used as a tool by the husband to extractresources from his wife.

The term α(I) denotes husband’s non-monetary utility from inflicting violence where whereα(I) > 0. It captures expressive violence in the framework. With the parameter, the benefit fromdomestic violence enters to the husband’s utility directly, not in an instrumental way. The parame-ter can be seen as the husband’s taste for violence and captures how household resources can affectthis taste. As an example, the husband’s taste for violence can be decreasing in household resourceswhich in turn affects violence in the household. This resembles an income effect on husband’s tastefor violence and thus the actual violence. The term α(I) can also capture the effect of financialstress on violence in the household. The husband can dislike financial stress and α(I) can capturestress relief from violence. In such a case, when there are more resources in the household, there isless financial stress and marginal utility from stress relief is lower. Thus violence decreases. Bothscenarios are plausible and allows to conceptualize the effect of an increase in household resources(either via the wife’s or the husband’s earnings) on violence in a non-extractive way. In order tocapture the aforementioned scenarios, I assume that α(I) is decreasing in both the husband’s orwife’s earnings, ∂α(I)

∂ Ih= ∂α(I)

∂ Iw< 0.

The term θ j for j ∈ {w,h} indicates the private level of satisfaction with the marriage for thewife and husband. As a key assumption of the model, the level of satisfaction with the marriageremains private information for each spouse. θ j follows distribution Fj with support [θ j,θ j]. θ j

is high enough for the marriage to be intact. vw is the disutility from violence for the wife wherevw > 0. Uw is decreasing in domestic violence and increasing in her and the husband’s earnings andher private level of satisfaction with the marriage. Uh is increasing in his and his wife’s earningsand his private level of satisfaction with the marriage.

34

I assume that their outside options can be represented by the utility functions

Sh = Ih and Sw = Iw. (5)

Each spouse only enjoys her/his personal earnings as an outside option. I define the outside optionas the utility of being separated. Due to paid employment, the wife has personal earnings tosupport herself financially if she initiates separation. I conceptualize the utility of being separatedwith personal earnings only as a simplification. Having a paid job also has non-monetary benefitsthat can constitute a part of the wife’s outside option. As an example, due to having a job, the wifehas a social network that can provide her a place to stay during the time she is separated.

Timing. The husband keeps a portion of his personal income to himself and gives the remainingportion to the wife for her expenses. The wife keeps a portion of her personal income and turnsthe rest to her husband. The husband observes the resources he receives from his wife, (1− x)Iw,and either chooses to inflict violence to extract more resources from the wife, τIw, or he choosesnot to be violent and remains with (1− x)Iw. If he chooses to inflict violence, his utility becomesUh = yIh +(1− x+ τ)Iw +αh(Ih + Iw)+θh. If he chooses not to, then Uh = yIh +(1− x)Iw +θh.Then, the wife either chooses to separate from the husband or not. If the husband chooses to inflictviolence and wife does not separate, her utility becomes Uw = (1−y)Ih+(x−τ)Iw+θw−vw wherevw > 0. In the absence of violence, her utility becomes Uw = (1− y)Ih + xIw + θw and she stayswith/does not separate from the husband assuming (1−y)Ih+xIw+θw > Sw. The husband’s utilityis high enough to stay with the wife in the absence of violence as well. When they are separated,the wife and the husband receive Sw and Sh respectively, their individual labor market potential.To highlight again, the level of satisfaction with the marriage remains private information for eachspouse throughout the marriage.

Decisions. The solution of the game between the husband and wife can be found via backwardinduction. The wife chooses between staying in the marriage and separating from the husbandgiven the husband’s decision to choose violence. She will separate from her husband if her utilityof staying in a violent marriage is smaller than the utility of being separated as in

(1− y)Ih +(x− τ)Iw +θw− vw︸︷︷︸Utility of being in a violent marriage

≤ Iw︸︷︷︸Utility of being separated

. (6)

Based on equation 6, there is a threshold value of the wife’s private level of satisfaction with themarriage, θw, that makes her indifferent between remaining together with and being separated from

35

her husband. The value of θw is

θw = (1− x+ τ)Iw− (1− y)Ih + vw. (7)

Recall that the cumulative distribution function of θw is given by Fw. Therefore, the probability ofthe wife separating from her husband is P(θw ≤ θw) = Fw(θw). Observe that as Iw increases, theprobability of separation increases. This is due to the fact that an increase in wife’s personal incomeis improving her outside option. Also, when Ih increases, the probability of separation decreases.This is because the wife enjoys a portion of the husband’s personal income if she remains married.

Since the level of satisfaction with the marriage is private information, the husband chooses toinflict violence only knowing the probability that she will initiate separation if her utility of stayingtogether is less than her utility of being separated. He compares the expected utility of choosingviolence, EUh, with utility of not choosing to be violent. His expected utility of inflicting violenceon his wife is

EUh = Fw(θw)Ih +[1−Fw(θw)][yIh +(1− x+ τ)Iw +α(I)+θh]. (8)

The husband chooses to be violent if

EUh︸︷︷︸Expected utility of inflicting violence

≥ yIh +(1− x)Iw +θh︸︷︷︸Utility from not inflicting violence

. (9)

Based on equation 9, there is a threshold value of the husband’s private level of satisfaction with themarriage, θh, that makes the husband indifferent between choosing violence (to extract a portionof the wife’s resources) and not being violent. The value of θh is determined by

θh = (1− y)Ih− (1− x+ τ)Iw−αh(Ih, Iw)+τIw +α(I)

Fw(θw). (10)

Therefore, the probability of the husband choosing violence is P(θh < θh) = Fh(θh).

7.2 Effect of a Mill Opening on Domestic Violence

For the couples who are exposed to a mill, the mill provides paid employment opportunities to thewife who was working as unpaid family workers. Thus, Iw increases. Since the husband receiveshigher earnings from selling cherries to the mill compared to selling home-processed coffee tomiddlemen in the domestic market, there is also an increase in Ih. I present the main channels thataffect domestic violence below.

36

Increase in Outside Option. As the wife’s earnings increases, her outside option improves. Thisincreases the probability of her initiating separation and makes it less likely for the husband toinflict violence on her.

Increased Incentives for Extractive Violence. As the wife’s earnings increases, the husband isincentivized to choose violence to extract the newly acquired resources from the wife.

Increase in Household Resources. As household resources increase either via the increase inhusband’s or women’s earnings, the husband would receive less non-monetary benefit from usingviolence. Thus he is less likely to inflict violence.

Differentiating the husband’s probability of choosing violence, Fh(θh), with respect to thewife’s earnings, ∂Fh(θh)

∂ Iw, highlights the impact of providing paid employment opportunities to

women on domestic violence. Differentiating it with respect to the husband’s earnings, ∂Fh(θh)∂ Ih

,highlights the impact of an increase in husband’s earnings on him inflicting violence. Both, impactof an increase in Iw and Ih on Fh(θh), are ambiguous.29 Under certain conditions, the effects maylead to a decline in violence.

Proposition. The effect of providing job opportunities to women on domestic violence, ∂Fh(θh)∂ Iw

, is

negative when Iw is large enough and τ and ∂α(I)∂ Iw

are small enough. The effect of an increase in

husband’s earnings on violence, ∂Fh(θh)∂ Ih

, is negative when ∂α(I)∂ Ih

is small enough.

Proof. See Appendix B.

The proposition highlights how the wife’s job opportunities may decrease domestic violence.When the wife’s outside option due to her job is good enough, the husband’s extraction rate issmall enough and he receives less marginal non-monetary benefit from violence with the wife’scontribution to household earnings, violence decreases. The proposition also highlights that keep-ing the wife’s earnings constant, an increase in the husband’s earnings may also decrease violence.This happens when the increase in his earnings (also an increase in household earnings), decreasehis non-monetary benefit from inflicting violence.30

29The effects still exist if decreasing marginal utility of consumption is assumed such as Uh = log(.) + θh andUw = log(.)+θw− vw. Results are available upon request.

30If such an increase in his earnings does not have any effect on his non-monetary benefit from violence, theviolence increases. Intuitively, this is because an increase in his earnings increases the benefit of marriage for the wife.Thus, she will less likely to leave when she experiences domestic violence. This makes the husband more likely toinflict violence on his wife.

37

8 Mechanisms behind the Decline in Domestic Violence

In this section, I disentangle the potential mechanisms behind the decline in domestic violenceempirically. To do so, I analyze the impact of a mill opening on bargaining power related out-comes, do a subsample analysis and analyze the seasonality of the decline in hospitalizations fordomestic violence within a year. As outlined in the literature and model sections, there can bemultiple reasons behind a decline in domestic violence. These include the increase in women’soutside options, decline in husbands’ non-monetary utility for domestic violence with an increasein household earnings (via an increase in women’s and/or husbands’ earnings) and exposure re-duction. While I cannot disentangle the exact mechanism, my analysis suggests that the declinein violence is driven by the channels that are active via women’s paid employment. Those chan-nels are the increase in women’s outside options and their contribution to household earnings, notexposure reduction. The increase in husbands’ earnings is not the dominant mechanism.

Increase in the Outside Option. According to the family economics literature, an increase inwomen’s outside option increases her bargaining power within the household. Using data on house-hold decision-making, I investigate whether women’s employment opportunities translate into animprovement in women’s bargaining power. Table 7 presents the results of estimating the impact ofmill exposure on household decision-decision making. Three different decisions are investigated.These are making large household purchases, making decisions on own health and visiting family.The estimation captures whether women are more likely to make these decisions by themselves orjointly with their husbands compared to their husbands or someone else in the family is making thedecision for them. When a within district approach is used, women who are exposed to a mill are5 percentage points (p-value= 0.02) more likely to make decisions on large household purchasesby themselves or jointly with their husbands. The estimated impact represents an increase of 7%with respect to the sample mean (0.69). There is no change in decision making related to women’sown health and visiting family. These results suggests that mill exposure improves women’s sayon financial decisions in the household.

I also investigate the impact of mill exposure on couple’s decision-making on contraception.Table 8 presents the results. The estimation captures whether women’s decision to use contracep-tion is taken jointly with their husbands compared to alone or husbands are making the decision forthem. For this variable, making the decision alone is not a proxy for bargaining power. Researchfinds that women in Sub-Saharan Africa may use contraception covertly as a response to their hus-bands’ desire to have more children (Ashraf et al., 2014). When within district approach is used,upon a mill opening, among women under the age of 40, women who are exposed to a mill are 5percentage points (p-value= 0.07) more likely to decide to use contraception jointly with their hus-

38

bands. The estimated impact represents an increase of 6% with respect to the sample mean (0.87).This is accompanied with a decline in the likelihood of women deciding to use contraception bythemselves. This suggests that the women who were making the contraception decision on theirown start to jointly decide to use contraception with their husbands. I do not have data on whetherwomen who decide to use contraception on their own are the ones who use contraception covertly.Yet, the increase in joint decision-making accompanied with a decline in women making theircontraception decision alone suggest that a mill opening increases communication and negotiationbetween couples.

Table 7: Effect of Mill Exposure on Household Decision Making: Woman Alone or Jointly withHusband


(1) (2) (3) (4) (5) (6)Large HHPurchases

OwnHealth

FamilyVisit

Large HHPurchases

OwnHealth

FamilyVisit

Mill 0.05** 0.02 0.02 0.03 -0.01 -0.00(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All dependent variables are coded as 1 ifthe decision is taken by women alone or jointly with their husbands and zero otherwise. Otherwisecategory includes husband only or someone else. The survey question on household decisions changedsignificantly over the data cycles. in 2005, the question asks “Who has the final say on decision x?” After2005, the question is “Who usually decides on x”. The answer option “Respondent and other” is alsodiscontinued in 2010 and 2014 rounds. To make a comparable variable, I create the variable based onwhether the woman had a say in the decision, either alone or jointly. Results should be interpreted withcaution. 4 km catchment area is used. *** p<.01, ** p<.05, * p<.1

Taken together, results on household decisions suggest that women’s improvement in theiroutside options due to mill exposure also improved their bargaining power, say, in some householddecisions. These results support the argument that an increase in women’s outside options due topaid employment is a potential mechanism behind the decline in domestic violence.

Increase in the Outside Option and Seasonality of Hospitalizations. No change in hospitaliza-tions during the post-harvest months relative to the hospitalizations in February (one month beforeharvest) further speaks to the outside option mechanism. After the harvest months, women areplausibly not able to credibly threat their husbands with separation. In a context where wages

39

are low, access to credit markets and saving technologies are limited, it is plausible for seasonaljobs to improve women’s outside options only when women have the jobs. First, schools reopenright after the harvest season, August. Although education is free in Rwanda upto nine years,parents have school expenses, primarily a school uniform fee. Anecdotal evidence suggests thatwomen buy school uniforms with the money they earn from the mills, so that their children cango to elementary school. In Table A.19, I report the results of estimating the impact of mill ex-posure on household’s education expenditures for elementary school-age children among coupleswho work in agriculture (school uniform expenses are included in the variable). 1% increase inmills per capita increases the education expenditures by 1.3%. If women are more likely to investin their children’s education compared to their husbands, they have less resources left to save forpost-harvest months.31 Second, the share of women who do not have a savings account is high inRwanda. 65% of the women in my sample do not have a savings account and this number increasesto 73% in rural areas. Limited access to savings accounts can affect how much women save andthis can affect women’s outside options during post-harvest months negatively.32

Table 8: Effect of Mill Exposure on (Modern) Contraception Decision-Making: Joint betweenCouples, Woman Only, Husband Only


(1) (2) (3) (4) (5) (6)Joint Woman Husband Joint Woman Husband

Mill 0.05* -0.04* -0.01 0.05 -0.03 -0.02(0.03) (0.03) (0.01) (0.03) (0.03) (0.01)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women younger than 40 who married before the expansion of the mills. 4 km catchment area isused.. *** p<.01, ** p<.05, * p<.1

Exposure Reduction. Another possible mechanism behind the decline in domestic violence canbe the reduction in the time couples spend together. Working hours in the mill are long. If womenwere farming in the same plot with their husbands before the expansion, starting to work in themills for pay is a shock to the time they are exposed to their husbands as well as their earnings.

31Doepke and Tertilt (2019) finds that in Mexico, cash transfers to women (PROGRESA) lead to an increase inspending on children and a decline in the savings rate.

32Anderson and Baland (2002) shows that in Kenya, participating in roscas, rotating savings and credit associations,is the wife’s strategy to protect her savings against claims by her husband for immediate consumption.

40

To provide evidence that exposure reduction is unlikely to drive my results, I take advantage ofthe information on women and their husbands’ occupation. I perform a subsample analysis. I runequation 1 using the sample of women who do not share the same occupation with their husbands.Self-employed women working in agriculture (smallholder farmers) whose husbands have skilledor unskilled manual jobs (truck drivers, plumbers, refuse workers, laborers in mining, construc-tion and manufacturing etc.) constitute the majority of the subsample. Given that those coupleshave different occupations, they do not work together during working hours. Thus, women startingworking in the mills will not be a shock to the time couples spend together. Their exposure toeach other will continue to be the same (no exposure during work hours). However, the workplaceof women will be shifted to mills where they work for pay. If there is a statistically significantincrease in the probability of women working for cash and a decline in the probability of experi-encing domestic violence in the past 12 months for this subsample, it suggests that my main result,reduction in domestic violence, is plausibly not driven by exposure reduction.

Results for women’s employment and domestic violence are reported in Table 9. Only withinthe district approach is used due to sample size. Upon a mill opening, being exposed to a millincreases the probability of working for cash in the past 12 months by 10 percentage points wherethe probability of working in the past 12 months remain unchanged. The estimated impact repre-sents an increase of 32% with respect to the sample mean (0.31). Moreover, being exposed to amill still decreases the probability of experiencing domestic violence in the past 12 months by 11percentage points. The estimated impact represents a decrease of 32% with respect to the samplemean (0.34). The results show that there is a statistically significant decline in the probability ofexperiencing domestic violence even among women whose exposure to their husbands are pre-sumably not affected by the expansion. Moreover, the magnitude of the decline is also similar(11 percentage points, 0.34 mean) to the result based on the whole sample (9 percentage points,0.35 mean). This further confirms that exposure reduction is not necessarily the main driver of thedecline in domestic violence.

One potential concern is whether the subsample is among couples with very educated husbandscompared to the main sample. This is because in the current subsample, husband occupations arenon-agricultural jobs. However, according to the data, the mean number of husbands’ education is5.7 for this subsample and it is 4.4 for the whole sample (husbands working in agriculture is notdropped). Taken together, the exercise provides suggestive evidence that exposure reduction is notthe mechanism behind my main results. I also further restricted the subsample to women who arefarmers and married to men with non-agricultural jobs. There is a statistically significant increasein working for cash for this subsample. I still find a negative estimate for the domestic violenceresult. However, due to low sample size, I do not have enough power to defend my (statistically

41

Table 9: Effect of Mill Exposure on Women’s Employment, Type of Earnings and Self-ReportedDomestic Violence in the Past 12 Months: Couples with Different Occupations

Within District

(1) (2) (3)Work Cash Violence

Mill -0.00 0.09** -0.11*(0.03) (0.04) (0.06)

Observations 3981 2893 1410Dependent variable mean 0.71 0.31 0.34

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. For the exposure reduction robustnesscheck, the sample is restricted to women who are not working with their husbands. The restriction is madebased on the occupation of women and their husbands. All dependent variables are measured for the past12 months. 4 km catchment area is used. *** p<.01, ** p<.05, * p<.1

insignificant) negative estimate.

Increase in Household Resources. A mill opening increases household earnings via women’swages as well as husbands receiving higher earnings from selling cherries to the mill comparedto selling home-processed coffee in the domestic market. Based on my conceptual framework,keeping women’s earnings constant, an increase in husbands’ earnings decreases their marginalnon-monetary benefit from inflicting violence. Intuitively, more resources in the household canmake the husband less financially stressed. If the husband inflicts violence for relieving finan-cial stress, an increase in his resources may decrease violence. Thus, the increase in husbands’earnings, not women’s paid employment, can be the main driver behind the decline in domesticviolence in the data. In order to investigate this mechanism, I build on the results from the previousexercise.

Recall that there is a decline in domestic violence even among couples with different occu-pations. Women working in agriculture and husbands working in non-agricultural manual jobsconstitute the majority of that subsample. To provide evidence that the increase in husbands’ earn-ings is unlikely to be the only mechanism, I estimate the impact of a mill opening on earnings ofwomen who work in agriculture and husbands who work in non-agricultural (again mostly man-ual) jobs. If there is no change in husbands’ earnings for this subsample, then there is a decline inviolence even among couples where husbands’ earnings do not increase with mill exposure. Thiswill suggest that the decline in husbands’ non-monetary benefit from inflicting violence due to an

42

increase in their earnings is not the only mechanism behind the decline in domestic violence.

Results are reported in Table 10. I find that mill exposure significantly increases women’sdaily earnings, however, there is no statistically significant change in their husbands’ earnings.The magnitude is also close to 0. This is plausible. Although mills may demand paid labor frommen with non-agricultural occupations (construction workers to build a mill, technicians to providerepairs in a mill), the expansion is primarily a (persistent) shock to individuals who are working inagriculture. No effect on husbands’ earnings shows that there is decline in violence even amongcouples where husbands’ earnings do not change with mill exposure. Moreover, the magnitude ofthe decline in violence is similar (11 percentage points, 0.34 mean) to the result based on the wholesample (9 percentage points, 0.35 mean). These evidence confirm that an increase in husbands’earnings is unlikely to be the main/only mechanism behind my results.

Table 10: Effect of a Mill Opening on Log of Last Daily Earnings: Couples with Different Occu-pations

Couples with Different Occupations

(1) (2)Woman’s Log of

Last Daily Earnings:Agriculture

Husband’s Log ofLast Daily Earnings:

Non-Agriculture

Log of Mills percapita in the District 2.65*** 0.97

(0.56) (1.00)

Observations 1828 1050Dependent variable mean 6.18 6.71

Note: Robust standard errors clustered at the district level are in parentheses. All estimations includeindividual controls, cohort fixed effects, district-by-year fixed effects, linear time trends interacted withbaseline district level characteristics. Sample consists of couples who married before the expansion of themills and have different occupations. Women report their occupation as agricultural where the husbandsreport their occupation as non-agricultural. Non-agricultural occupations mostly consist of manual(unskilled and skilled) jobs. Since EICV is not geocoded, the mill variable is log of total number of millsper capita in the district in a given year. *** p<.01, ** p<.05, * p<.1

Recall from the model that the husband’s marginal non-monetary benefit can also decrease dueto an increase in women’s earnings. For this sample, there is an increase in women’s earnings thatis accompanied with no change in husbands’ earnings and there is a decline in domestic violence.This suggests that women’s contribution to household earnings is also a plausible mechanism be-hind the decline in violence.

It should also be noted that the increase in women’s earnings for these couples is bigger com-

43

pared to the increase for the overall sample where majority of the couples are both farmers. Thereason is as follows. Women who are married to men with non-agricultural manual jobs are plausi-bly earning income from both selling coffee cherries to the mill and working in the mill as a wageworker (or earns from selling cherries only which still means earning a higher amount of moneycompared to working in the mill only). This confirms that among the farmer couples, money earnedby selling coffee is husbands’ money.

It is also valid to ask what were women who are married to non-farmer men were doing beforethe expansion. According to DHS data, in 2004 and 2005, before the expansion, 59% of thosewomen reported agricultural self-employed as their occupation. 61% of those self-employed areunpaid workers before the expansion. This suggests that they were subsistence farmers that didnot use cash-crops. To understand the effect of a mill opening on them better, I use information oncrops and find that the share of farmer women who have coffee in their family plots increases withmill expansion over the years but the share of farmer women who have beans in their plots do notchange over time. This suggests that those women were not having coffee on their plots before,but they planted coffee trees with the expansion.

Increase in Household Earnings and Seasonality of Hospitalizations. No change in hospitaliza-tions during post-harvest, let’s say August, relative to the hospitalizations in February (one monthbefore the harvest) further speaks to this mechanism. In a context where household resourcesare low, access to credit markets and saving technologies are limited, the effects of providing jobopportunities to women may be concentrated only during the period when women have the jobs.After the harvest months, the household has less resources since women do not work for pay. Thismay increase the husband’s financial stress and lead to an increase in domestic violence. Accordingto EICV data (earnings data), in 2010, among couples where women work in agriculture and havepositive earnings, women’s last daily earnings constitute 38% of their households’ daily food con-sumption expenditures.33 A descriptive study from Nyamasheke district, Bayisenge et al. (2019),shows that households benefit women’s earnings from coffee production. The study documentsthat in the district, most of coffee farmer women’s earnings (68.75%) are spent to household dailyneeds. If women contribute to household consumption with their wages from the mills during theharvest months, the discontinuity of her contribution may lead to an increase in financial stress andthus violence. This may potentially be one of the reasons why the hospitalizations in the catchmentareas shift back to its pre-harvest level, when women do not work for pay.34

33Husbands’ last daily earnings constitute for 55% of the households’ daily food expenditures34Recall that the husbands may receive a second payment from the mill in November. Thus in November, there

may be an increase in household earnings via husbands only and women do not work for pay. In November, thereis a statistically insignificant decline in hospitalizations and the magnitude is half of the effect during the peak of theharvest. This is in line with the theoretical prediction that the larger the women’s outside options and their contributionto household earnings, the larger the effect is in magnitude. When both channels are not present, a statistically

44

Mechanisms and Women’s Paid Employment. My analysis suggests that the decline in violenceis driven by the channels that are active via women’s paid employment. Those channels are theincrease in women’s outside options and women’s contribution to household earnings. An increasein husbands’ earnings is not the dominant mechanism behind the decline in domestic violence.Moreover, I rule out exposure reduction, that is not depicted in the model, yet documented in theliterature.

The result that there is a decline in hospitalizations for domestic violence only when womenare working for pay reinforces that the decline in violence is plausibly driven by women’s paidemployment. It further confirms the increase in women’s outside options and their contributionto household earnings as mechanisms. It also suggests that in a context where wages are low,household resources, access to credit markets and saving technologies are limited, the effect isconcentrated when women have the jobs.

Research suggests that employment has non-monetary benefits (Hussam et al., 2021). In thiscontext, non-monetary value of earning personal income (confidence, dignity etc.) and having asocial network (coworkers in the mills) are such potential benefits. Although they plausibly affectthe decline in domestic violence, unfortunately I am not able to quantify them with my data. Yet,both channels are active via women’s paid work and I show that the decline in domestic violence inmy context is plausibly driven by women’s paid employment. As future research, I aim to furtherdisentangle how paid job opportunities affect domestic violence.

9 Robustness Checks

9.1 Dynamic Impact of Mill Exposure and Pre-Trends

I also estimate the impact of a mill opening using an even-study specification. The specificationexploits the variation in years when individuals are exposed to a mill. The results provide insightson the dynamics of the effects, the impact of a mill in the short and long-run. Moreover, estimatesduring the years before mill exposure constitute a test for the parallel trends assumption. I estimatethe dynamic impact of a mill opening with the event study specification below:

Yistk = β0 +6

∑k=−3

Millis×βk1[τ = k]+Xistφ +λc +ωm +αs + γdt +(Xs× t)θ + εistk. (11)

significant prominent decline in domestic violence is not observed. Lastly, increase in husbands’ earnings that isaccompanied with no increase in domestic violence hospitalizations in November also suggests that husbands’ non-monetary utility for domestic violence is not zero and plausibly decreasing in household earnings. Yet, this discussionshould be reviewed with caution since there is limited knowledge on receiving second payments in the context.

45

The dependent variable Yistk is the outcome of interest of woman i (or the husband of woman i), insector s, in year t and at event-time k. Millsis is a binary variable coded as 1 if woman/husband i

resides within the catchment area of a mill and zero otherwise by the end of the sample, 2014. It isinteracted with event-year dummies, 1[τ = k], to investigate the dynamic impact of a mill opening.τ denotes the event-year. τ = 0 represents mill’s year of operation. For k > 0, τ = k represents k

years after a mill’s opening. For k < 0, τ = k represents k years before a mill’s year of operation.The omitted category is τ =−1, which means that the dynamic impact of being exposed to a millis estimated with respect to one year prior to a mill opening. Xist is the same set of controls inthe static empirical specification. λc is the cohort fixed effects. ωm is the year of marriage fixedeffects. αs is the sector fixed effects. γdt is the district-by-year fixed effects. Xs is a vector ofbaseline geographical variables interacted with linear time trends. I clustered standard errors at thedistrict level. The main dependent variables for this specification are being employed, working forcash and experiencing domestic violence in the past 12 months.

The coefficients of interest are βks when k ≥ 0. For k ≥ 0, each βk provides the change inoutcomes for the individuals who are exposed to a mill relative to the individuals who are notwithin the same district, k years after a mill opening, relative to one year before a mill’s yearof operation. Since k < 0 represents the years before a mill opening, I expect to see small andstatistically insignificant estimates for such βks when k < 0 . If so, such estimates will constitutesupportive evidence in favor of the parallel trends assumption.

Figure 7 presents the results from estimating equation 11 (event-study specification). It plotsthe coefficient of the interaction term for 3 years before and 6 years after a mill opening. I use thewithin district approach to have balanced number of observations every time period. The dynamicresults are mostly in line with the static estimates. I find a statistically significant decrease in theprobability of experiencing domestic violence in the past 12 months right after a mill opening.Exposure to a mill opening also increases women’s probability of working for cash starting fromright after a mill opening. There is no change in women’s probability of working. I need tonote that number of women who are in the treatment group is lower for the event times 1, 2 and3 compared to the remaining event times. This is because the number of mill openings are notbalanced across the 2000s and I have the DHS data cycles for specific years. This can plausiblybe the reason why the coefficients at event times 1, 2 and 3 are statistically insignificant for theworking for cash variable although they are positive.

I should also note that there is no statistically significant decline in DV at event times 4, 5 and6, although women are more likely to work for cash in these years relative to one year before a millopening. The number of observations is not low compared to other years for these event times.This suggests that the effect of women’s paid employment in low-paid jobs on domestic violence

46

may not persist for long years.

Figure 7: Dynamic Impact of a Mill Opening on Self-Reported Domestic Violence, Women’s Typeof Earnings and Employment in the Past 12 Months

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All variables are measured for the past 12months. 4 km catchment area is used. Within district approach is used. *** p<.01, ** p<.05, * p<.1

47

Figure 8 plots the coefficient of the interaction terms for the same specification for husbands.As expected, there is no change in probability of working and working for cash in the past 12months right after a mill opening. I need to note that there is a statistically significant declinein the husband’s probability of working only at event time 2. For working for cash variable, theestimates for event times 1 and 2 are also much bigger than 0 compared to the remaining eventtimes, although insignificant. This is plausibly due to the lower number of observations at eventtimes 1, 2 and 3 based on my data and the imbalance in the mill expansion.

Figure 8: Dynamic Impact of a Mill Opening on Husbands’ Employment and Type of Earnings inthe Past 12 Months

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartners of women in the first specification. All variables are measured for the past 12 months. 4 kmcatchment area is used. *** p<.01, ** p<.05, * p<.1

For all variables, the coefficients are close to zero and statistically insignificant for the yearsbefore a mill opening. Thus, the estimates constitute evidence in favor of the parallel trends as-sumption. All results are also shown in Tables A.23, A.24, A.25 and A.26.

9.2 Role of Variation in Treatment Timing

Goodman-Bacon (2021) demonstrates that twoway fixed effects difference-in-differences (TWFEDD)estimator is a weighted average of all possible 2x2 DD estimators that compare timing groups toeach other. Under time-varying treatment effects, TWFEDD estimates a weighted average treat-ment effect with negative weights (Borusyak and Jaravel 2018, de Chaisemartin and D’Haultfœuille2020, Sun and Abraham 2020, Callaway and Sant’Anna 2020, Goodman-Bacon 2021) which bi-ases the results. This occurs when the later-treated group uses the earlier-treated group as control.

48

Goodman-Bacon (2021) highlights that specifically the units treated near the beginning or the endof the panel can get more weight as controls than treatments, which is always the case in designswithout untreated units. I do not expect this to be a major concern in my context. I have untreatedunits and out of the 211 mill openings as of 2002, only 21 percent of them opened near the be-ginning or the end of the panel (2012). 79% of the mills opened between 2005-2010. In order toanalyze the 2x2 DD comparisons and weights formally, I perform the Goodman-Bacon (2021) de-composition.35 The decomposition calculates the weights of each type of 2x2 DD comparison andhow it contributes to the TWFEDD estimate. I find that 80-81% of the TWFEDD estimate for labormarket outcomes (working for pay, working cash for the past 12 months. and domestic violenceis derived from Treatment vs. Never Treated comparisons. Earlier Group Treated vs. Later GroupControl and Later Group Treatment Control each has 0.01 weights and do not contribute much tothe TWFEDD estimate as expected. The weight for Treated vs. Already Treated comparison isonly 0.16.

I also follow the test proposed by de Chaisemartin and D’Haultfœuille (2020) which enablesresearchers to understand when treatment effect heterogeneity would be a serious concern for thevalidity of their estimates.36 de Chaisemartin and D’Haultfœuille (2020) recommends researchersto calculate both the weights and the ratio of the TWFE estimate divided by the standard deviationof the weights. If many weights are negative, and if the ratio is not very large, then treatment effectheterogeneity would be a serious concern for the validity of the estimate.37 For the working for payand cash in the past 12 months outcomes, I find that 92 and 93% of my treatment effects receivepositive weights respectively where the remaining receive negative weights. I also calculate theratios based on my estimates and the standard errors of the weights and they are sufficiently high.Thus, given that approximately only 7% of the weights are negative and the ratios are sufficientlyhigh, treatment effect heterogeneity is not a concern for the validity of my estimates. For a com-plete robustness check, I estimate my results using the estimator proposed in de Chaisemartin andD’Haultfœuille (2020) that gives valid results even if the treatment effect is heterogeneous overtime and across groups. I plot the estimation results for main results in Figure A.6.38 As expected,the estimates are very similar to my main results.39

Sun and Abraham (2020) demonstrates that in event study designs with variation in treatment

35I implement the decomposition following Goodman-Bacon (2021) and based on bacondecomp STATA package.36My calculations are based on twowayfeweights STATA package.37If that ratio is close to 0, that estimate and the average treatment effect can be of opposite signs even under a

small and plausible amount of treatment effect heterogeneity. This is the case where, treatment effect heterogeneitywould be a serious concern for the validity of that estimate.

38The results are based on didmultiplegt STATA package.39The weights and the results in Figure 10 are based on within district approach. The results are also robust and

similar when donut approach is used.

49

timing across units, the coefficient on a given lag or lead can be contaminated from other periods.Pretrends can arise from heterogeneity of the treatment effects as well. As a robustness check formy event study results, I perform the alternative estimation method proposed in Sun and Abraham(2020) that is free from such contamination.40 The event study results from this method is reportedin Figure A.7 and very similar to my main event study results.

9.3 Measuring the Catchment Area

9.3.1 Using Different Buffer Radii for the Catchment Area

Throughout the paper, I use a 4 km radius buffer around a mill to construct the treatment group.I also experiment with multiple radii to construct the treatment group using within the districtapproach for the control group. I show that as the radius of the catchment area increases, meaningmore DHS clusters who are actually not treated become a part of the treatment group, my resultsfade out. Tables A.27, A.28 and Figures A.9, A.10 presents and plots the results using DHS andHMIS (hospital) data. This rules out the concern that the households who are right outside of thecatchment area are sorting themselves into the treatment group.

If the coffee farmers who reside outside of the catchment areas bring their coffee cherries to themills, the cherries will rot by the time they arrive to the mills. Thus, the husbands will not receiveearnings from selling cherries to the mills. Yet, women can walk all the distance and request a jobin the mills since mills do not prohibit those women to work there as wage workers. There aretwo potential reasons why I do not observe this in my data. First, the paid jobs in the mills are inshort supply. There are not enough jobs for both the coffee farmer women who reside within thecatchment area as well as the women who reside outside. However, it is also possible that womenwho reside within the catchment areas can hire women who are just outside to work for them inthe mills.

Suppose women in the catchment areas may work in the mills for three days during the weekand hire the women outside on the fourth day with the money they earned from working previously.If that scenario happens, women outside of the catchment areas will be treated by the mills. Thisbrings us to the second potential reason why I do not observe this in my data. It is likely thatthe wage in the mills is lower than the reservation wage of the woman outside of the catchmentarea. The woman farmer who is outside is processing coffee at home and the husband sells theprocessed coffee in the domestic market. Thus, she contributes to agricultural production in thehousehold. She also contributes to home production by performing subsistence farming (wheat

40The results are based on eventstudyinteract STATA package.

50

for bread etc.) for household consumption. It is plausible that the wage in the mills is not highenough to compensate the total of the money she receives from her husband for her contribution tothe household’s agricultural production and her contribution to the home production. In contrast,a household in the catchment area can afford to buy food from the market rather than the womanproducing it via home production every day. This is because the husband has higher earnings fromselling cherries to the mill compared to selling home-processed coffee in the local market. Thisis another way of saying that in the catchment area, the household income without women’s paidwork is high enough (plausibly due to husbands’ earnings from the mills) for the women to notperform subsistence farming every day, do work in the mills for long hours and bring more moneyto the household.

This robustness check also alleviates the concern regarding the random displacement of DHSclusters and measurement error. The results are weakly significant when 5 km buffer radius is usedand there is no effect when the radius is 10 km. As the buffer radius increases, there is a gradualfading of the results. This suggests that the 4 km buffer radius is a valid measure for the catchmentarea and I am not picking up an effect only due to the displacement of DHS clusters. Yet, it is stillpossible for the results to have measurement error when I use 4 km as the buffer radius. This isbecause the rural clusters contain a maximum of 5 kilometers of positional error, which is higherthan than the 4 km buffer radius. A DHS cluster that is originally not in the catchment area can be inthe catchment area due to the random displacement. Yet, I am still picking up an effect on women’spaid employment and domestic violence with a 4 km buffer radius and randomly displaced DHSclusters. Thus, the results constitute a lower bound for the effect and if I have precise coordinatesfor the clusters I would likely observe larger effects.

10 Conclusion

This paper investigates whether providing paid employment opportunities to women decrease theviolence they face from their partners using the transformation of the Rwandan coffee industryas a natural experiment. I answer my research question in two steps. First, I provide causal ev-idence that the government-induced rapid expansion of the coffee mills in Rwanda in the 2000s,increased women’s paid employment, earnings of women and their husbands and decreased do-mestic violence. To identification, I uniquely perform two strategies with two sources of domesticviolence data, self-reports and novel administrative data on the universe of monthly hospitaliza-tions. Then, I provide evidence to show that the decline in violence is plausibly driven by women’spaid employment, not by an increase in husbands’ earnings. First, after a mill opening, womenin the catchment areas are more likely to participate in household decisions. Second, there is a

51

decline in violence even among couples where women work in agriculture and husbands work inoccupations with no change in earnings with a mill, non-agricultural manual jobs. Since a millopening is not a shock to these couples’ shared time during work hours, this also suggests thatwomen’s paid employment does not affect violence via exposure reduction. The seasonality of thedecline in hospitalizations (decline in violence only during the harvest months when mills do notoperate, not before or after the season) further speaks to the mechanisms and reinforces that thedecline in violence is plausibly driven by women’s paid employment. When women can crediblythreat men with short term separation and when there is less financial stress in the household due towomen’s paid employment during the harvest months, domestic violence decreases. When womenstop earning personal incomes, domestic violence shifts back to its pre-harvest level.

My results provide insights to policymakers that aim to combat domestic violence. Based onmy results, providing job opportunities to women has the potential to alleviate domestic violencein developing countries. Sub-Saharan Africa is currently under a large-scale economic transitionthat provides paid job opportunities to women. Although the region has the highest FLFP ratesglobally since the 1990s, more than 60% of the employed women are often unpaid agriculturalfamily workers in their family plots (FAO, 2010). Recently, between 2000 and 2018, 9 millionpaid jobs are created per year, mostly in the agricultural sector where women work dominantly(IMF, 2018). My results suggest that the transition has the potential to decrease the violencewomen face from their partners in the region.

52

References

AgriLogic (2018). Value Chain Analysis for the Coffee Sector in Rwanda. Technical Report July.

Aizer, A. (2010). The gender wage gap and domestic violence. American Economic Review 100(4),1847–1859.

AlAzzawi, S. (2014, dec). Trade liberalization, industry concentration and female workers: thecase of Egypt. IZA Journal of Labor Policy 3(1).

Alesina, A., B. Brioschi, and E. La Ferrara (2020). Violence Against Women: A Cross-culturalAnalysis for Africa. Economica.

Alesina, A., P. Giuliano, and N. Nunn (2013). On the origins of gender roles: Women and theplough. Quarterly Journal of Economics 128(2), 469–530.

Anderberg, D., H. Rainer, J. Wadsworth, and T. Wilson (2016). Unemployment and DomesticViolence: Theory and Evidence. Economic Journal 126(597), 1947–1979.

Anderson, S. and J. M. Baland (2002). The economics of roscas and intrahousehold resourceallocation. Quarterly Journal of Economics 117(3).

Anderson, S. and M. Eswaran (2009, nov). What determines female autonomy? Evidence fromBangladesh. Journal of Development Economics 90(2), 179–191.

Anderson, S. and G. Genicot (2015). Suicide and property rights in India. Journal of Development

Economics 114, 64–78.

Angelucci, M. (2008). Love on the rocks: Domestic violence and alcohol abuse in rural Mexico.B.E. Journal of Economic Analysis and Policy 8(1).

Angelucci, M., D. Karlan, and J. Zinman (2015). Microcredit impacts: Evidence from a Random-ized Microcredit Program Placement Experiment by Compartamos Banco. American Economic

Journal: Applied Economics 7(1), 151–182.

Arenas-Arroyo, E., D. Fernandez-Kranz, and N. Nollenberger (2021, feb). Intimate partner vio-lence under forced cohabitation and economic stress: Evidence from the COVID-19 pandemic.Journal of Public Economics 194.

Ashraf, N., E. Field, and J. Lee (2014). Household bargaining and excess fertility: An experimentalstudy in zambia. American Economic Review 104(7), 2210–2237.

53

Baird, S., C. McIntosh, and B. Ozler (2019, sep). When the money runs out: Do cash transfershave sustained effects on human capital accumulation? Journal of Development Economics 140,169–185.

Banerjee, A., D. Karlan, and J. Zinman (2015). Six randomized evaluations of microcredit: Intro-duction and further steps. American Economic Journal: Applied Economics 7(1), 1–21.

Bayisenge, J. (2010). Access to Paid Work and Women’s Empowerment in Rwanda. 19(1),84–106.

Bayisenge, R., P. H. S. Shengde, Y. Harimana, J. B. Karega, M. Nasrullah, and D. Tuyiringire(2019). Gender Equality, Agriculture and Rural Development: Evidence from NyamashekeCoffee Production in Rwanda. International Journal of Gender and Women’s Studies 7(1).

Bhalotra, S., D. G. Britto, P. Pinotti, and B. Sampaio (2021). Job Displacement, UnemploymentBenefits and Domestic Violence. SSRN.

Bhalotra, S., U. Kambhampati, S. Rawlings, and Z. Siddique (2019). Intimate Partner Violence:The Influence of Job Opportunities for Men and Women. The World Bank Economic Review.

Blau, F. D. and L. M. Kahn (2017, sep). The gender wage gap: Extent, trends, & explanations.Journal of Economic Literature 55(3), 789–865.

Bloch, F. and V. Rao (2002). Terror as a bargaining instrument: A case study of dowry violence inrural India. American Economic Review 92(4), 1029–1043.

Bobonis, G. J., M. Gonzalez-Brenes, and R. Castro (2013). Public transfers and domestic violence:The roles of private information and spousal control. American Economic Journal: Economic

Policy 5(1), 179–205.

Borusyak, K. and X. Jaravel (2018, mar). Revisiting Event Study Designs. SSRN Electronic

Journal.

Boudreaux, K. C. (2011). Economic liberalization in Rwanda’s coffee sector: A better brew forsuccess. In Yes Africa Can, pp. 185.

Bursztyn, L., A. L. Gonzalez, and D. Yanagizawa-Drott (2020, oct). Misperceived social norms:women working outside the home in saudi arabia. American Economic Review 110(10), 2997–3029.

Callaway, B. and P. H. Sant’Anna (2020). Difference-in-Differences with multiple time periods.Journal of Econometrics.

54

Calvi, R. and A. Keskar (2021). ’Til Dowry Do Us Part: Bargaining and Violence in IndianFamilies.

Chiappori, P. A. and M. Mazzocco (2017). Static & intertemporal household decisions. Journal of

Economic Literature 55(3), 985–1045.

Chin, Y. M. (2011, dec). Male backlash, bargaining, or exposure reduction?: Women’s workingstatus and physical spousal violence in India. Journal of Population Economics 25(1), 175–200.

Church, R. A. (2018). What Transparent Trade Looks Like.

de Chaisemartin, C. and X. D’Haultfœuille (2020, sep). Two-Way Fixed Effects Estimators withHeterogeneous Treatment Effects. American Economic Review 110(9), 2964–2996.

Doepke, M. and M. Tertilt (2019). Does female empowerment promote economic development?Journal of Economic Growth 24(4).

Dugan, L., D. S. Nagin, and R. Rosenfeld (1999). Explaining the decline in intimate partner homi-cide: The effects of changing domesticity, women’s status, and domestic violence resources.Homicide Studies 3(3), 187–214.

Erten, B. and P. Keskin (2018). For better or for worse? Education and the prevalence of domesticviolence in Turkey. American Economic Journal: Applied Economics 10(1), 64–105.

Erten, B. and P. Keskin (2020). Breaking the Cycle? Education and the Intergenerational Trans-mission of Violence. The Review of Economics and Statistics 102(2), 252–268.

Erten, B. and P. Keskin (2021). No Trade-offs? The Impact of WTO Accession on Intimate PartnerViolence in Cambodia.

Eswaran, M. and N. Malhotra (2011). Domestic violence and women’s autonomy in devel-oping countries: theory and evidence. Canadian Journal of Economics/Revue canadienne

d’economique 44(4), 1222–1263.

FAO (2010). Women’s work. Technical report.

Farmer, A. and J. Tiefenthaler (1997). An economic analysis of domestic violence. Review of

Social Economy 55(3), 337–358.

Field, E., R. Pande, N. Rigol, S. Schaner, and C. T. Moore (2021, jul). On her own account: Howstrengthening women’s financial control impacts labor supply and gender norms. American

Economic Review 111(7), 2342–2375.

55

Gaddis, I. and J. Pieters (2017, mar). The Gendered Labor Market Impacts of Trade LiberalizationEvidence from Brazil. Journal of Human Resources 52(2), 457–490.

Gelles, R. (1974). The Violent Home: A Study of Physical Aggression Between Husbands and

Wives. Sage Publications.

Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. Journal

of Econometrics.

Guarnieri, E. and H. Rainer (2021, jun). Colonialism and female empowerment: A two-sidedlegacy. Journal of Development Economics 151.

Haushofer, J., C. Ringdal, J. P. Shapiro, and X. Y. Wang (2019). Income Changes and IntimatePartner Violence: Evidence from Unconditional Cash Transfers in Kenya. NBER Working paper.

Heath, R. (2014, may). Women’s Access to Labor Market Opportunities, Control of HouseholdResources, and Domestic Violence: Evidence from Bangladesh. World Development 57, 32–46.

Heath, R., M. Hidrobo, and S. Roy (2020, mar). Cash transfers, polygamy, and intimate partnerviolence: Experimental evidence from Mali. Journal of Development Economics 143.

Heath, R. and S. Jayachandran (2017, jan). The causes and consequences of increased femaleeducation and labor force participation in developing countries. In The Oxford Handbook of

Women and the Economy, pp. 345–367. Oxford University Press.

Heath, R. and M. Mobarak (2015, jul). Manufacturing growth and the lives of Bangladeshi women.Journal of Development Economics 115, 1–15.

Hidrobo, M., A. Peterman, and L. Heise (2016). The effect of cash, vouchers, and food transferson intimate partner violence: Evidence from a randomized experiment in Northern Ecuador.American Economic Journal: Applied Economics 8(3), 284–303.

Hussam, R., E. M. Kelley, G. Lane, and F. Zahra (2021). The Psychosocial Value of Employment.

Hyland, M., S. Djankov, and P. K. Goldberg (2020, dec). Gendered Laws and Women in theWorkforce. American Economic Review: Insights 2(4), 475–490.

ICRW (2011). Evolving Men: Gender Equality Survey. Technical report.

IMF (2018). The Future of Work in Sub-Saharan Africa. Technical Report 18.

56

Jensen, R. (2012, may). Do labor market opportunities affect young women’s work and familydecisions? Experimental evidence from India. Quarterly Journal of Economics 127(2), 753–792.

Keats, A. (2018, nov). Women’s schooling, fertility, and child health outcomes: Evidence fromUganda’s free primary education program. Journal of Development Economics 135, 142–159.

Khanna, M. and D. Pandey (2020). Reinforcing Gender Norms or Easing Housework Burdens?The Role of Mothers-in-Law in Determining Women’s Labor Force Participation.

Kruk, M. R., W. Meelis, J. Halasz, and J. Haller (2004, oct). Fast positive feedback between theadrenocortical stress response and a brain mechanism involved in aggressive behavior. Behav-

ioral Neuroscience 118(5), 1062–1070.

La Mattina, G. (2017). Civil conflict, domestic violence and intra-household bargaining in post-genocide Rwanda. Journal of Development Economics 124, 168–198.

Luke, N. and K. Munshi (2011, jan). Women as agents of change: Female income and mobility inIndia. Journal of Development Economics 94(1), 1–17.

Lundberg, S. and R. A. Pollak (1993, dec). Separate Spheres Bargaining and the Marriage Market.Journal of Political Economy 101(6), 988–1010.

Macchiavello, R. and A. Morjaria (2020). Competition and Relational Contracts in the RwandaCoffee Chain. The Quarterly Journal of Economics.

Macmillan, R. and R. Gartner (1999, nov). When She Brings Home the Bacon: Labor-ForceParticipation and the Risk of Spousal Violence against Women. Journal of Marriage and the

Family 61(4), 947.

Majlesi, K. (2016, mar). Labor market opportunities and women’s decision making power withinhouseholds. Journal of Development Economics 119, 34–47.

Manser, M. and M. Brown (1980). Marriage and Household Decision-Making: A BargainingAnalysis. International Economic Review 21(1), 31.

McElroy, M. B. and M. J. Horney (1981). Nash-Bargained Household Decisions: Toward a Gen-eralization of the Theory of Demand. International Economic Review 22(2), 333–349.

McKelway, M. (2021). Women’s Employment in India: Intra-Household and Intra-Personal Con-straints.

57

MINAGRI and MINICOM (2008). Rwanda National Coffee Strategy. Technical report.

Mulligan, C. B. and Y. Rubinstein (2008, aug). Selection, investment, and women’s relative wagesover time. Quarterly Journal of Economics 123(3), 1061–1110.

Perez-Heydrich, C., J. L. Warren, C. R. Burget, and M. E. Emch (2013). Guidelines on the Use ofDHS GPS Data. Spatial Analysis Reports No. 8. Technical report.

Qian, N. (2008, aug). Missing women and the price of tea in China: The effect of sex-specificearnings on sex imbalance. Quarterly Journal of Economics 123(3), 1251–1285.

Rogall, T. (2021, jan). Mobilizing the masses for genocide. American Economic Review 111(1),41–72.

Sanin, D. (2021). Do Domestic Violence Laws Protect Women From Domestic Violence ? Evi-dence From Rwanda.

SCAA Sustainability Council (2015). The Blueprint for Gender Equality in the Coffeelands. Tech-nical report.

Schilling, M. and T. McConnell (2004). Model for Siting Coffee Washing Stations in Rwanda.

Scholer, M. (2008). Women In Coffee. International Trade Forum (3/4), 32–33.

Sivasankaran, A. (2014). Work and Women s Marriage , Fertility and Empowerment : Evidencefrom Textile Mill Employment.

Sun, L. and S. Abraham (2020). Estimating dynamic treatment effects in event studies with het-erogeneous treatment effects. Journal of Econometrics.

Sviatschi, M. M. and I. Trako (2021). Gender Violence, Enforcement, and Human Capital: Evi-dence from Women’s Justice Centers in Peru.

Tauchen, H. V., A. D. Witte, and S. K. Long (1991). Domestic Violence: A Nonrandom Affair.International Economic Review, 491–511.

Tur-Prats, A. (2021, may). Unemployment and intimate partner violence: A Cultural approach.Journal of Economic Behavior and Organization 185, 27–49.

WHO (2013). Global and regional estimates of violence against women: Prevalence and healtheffects of intimate partner violence and non-partner sexual violence. Technical report.

58

World Bank (2015). Women, Business and the Law 2016: Getting to Equal. Technical report,World Bank.

Ya-Bititi, G. M., P. Lebailly, D. S. Mbonyinkebe, and P. Burny (2019, aug). ’Coffee has givenus power to act’: Coffee cooperatives and women’s empowerment in Rwanda’s rural areas: Acase study of Karaba coffee cooperative. In Cooperatives and the World of Work, pp. 107–118.Taylor and Francis.

Yucesoy, B., L. E. Charles, B. Baker, and C. M. Burchfiel (2015). Occupational and genetic riskfactors for osteoarthritis: A review. Work 50(2), 261–273.

59

Appendix

A Additional Tables and Figures

Table A.1: Summary Statistics of Mills

Mean Std. Dev.

Panel A: Mill Level Characteristics

Owner: Cooperative 0.5 0.5

Owner: NGO 0.2 0.4

Opened before 2005 0.1 0.3

Opened between 2005-2010 (Rapid Expansion period) 0.8 0.4

Panel B: Sector Level Characteristics

Number of coffee farmers in 1999 702.2 464.2

Log Coffee Trees in 1999 11.5 0.6




FAO-GAEZ Coffee Suitability Index for Coffee: Marginal 0.3 0.5

FAO-GAEZ Coffee Suitability Index for Coffee: Moderate 0.4 0.5

FAO-GAEZ Coffee Suitability Index for Coffee: Medium 0.2 0.4

FAO-GAEZ Coffee Suitability Index for Coffee: Good 0.0 0.2

Panel A: Spatial Characteristics

Sector Area in km2 52.4 26.3

District Area in km2 834.2 274.7

Note: Mill level characteristics are from the Rwanda GeoPortal and Macchiavello and Morjaria (2020).Data on coffee trees and farmers are from the coffee censuses. FAO-GAEZ Suitability Index categories arebased on FAO-GAEZ’s suitability categories. Aggregate crop suitability is (1) Very high when index >8500; (2) High when 7000 < index ≤ 8500; (3) Good when 5500 < index ≤ 7000; (4) Medium when 4000< index ≤ 5500; (5) Moderate when 2500 < index ≤ 4000; (6) Marginal when 1000 < index≤ 2500; (7)Very marginal when 0 < index ≤ 1000; and (8) Not suitable when index = 0. The continuous index is at 9km2 resolution. I aggregate the index at the sector level and classify the sector level index based on FAO’ssuitability categories. Spatial characteristics are the author’s calculations using spatial maps by the RwandaGeoPortal via ArcGIS software.

60

Tabl

eA

.2:S

umm

ary

Stat

istic

sfo

rWom

en:D

HS

Wom

en’s

Rec

ode

All

Not

Exp

osed

toa

mill

Exp

osed

toa

mill

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esW

orke

din

the

past

12m

onth

s0.

880.

320.

880.

320.

880.

32W

orke

dfo

rcas

hin

the

past

12m

onth

s0.

390.

490.

370.

480.

470.

50E

xper

ienc

eddo

mes

ticvi

olen

cein

the

past

12m

onth

s0.

340.

470.

330.

470.

370.

48

Pane

lB:C

ontr

ols

Hus

band

lives

inth

eho

use

0.88

0.32

0.89

0.32

0.87

0.34

Hus

band

’sag

e42

.42

9.85

42.2

09.

8843

.11

9.72

Hus

band

’sO

ccup

atio

n:A

gric

ultu

ral

0.72

0.45

0.71

0.45

0.73

0.44

Hus

band

’sed

ucat

ion

inye

ars

4.32

3.76

4.32

3.81

4.32

3.60

Occ

upat

ion:

Agr

icul

tura

l0.

770.

420.

760.

430.

790.

41M

arita

lsta

tus:

Mar

ried

0.73

0.44

0.71

0.45

0.80

0.40

Mon

ogam

y(N

oot

herw

ives

)0.

890.

310.

880.

330.

920.

27N

umbe

rofu

nion

s:O

ne0.

840.

370.

840.

370.

850.

36A

geat

first

mar

riag

e19

.84

3.33

19.7

43.

3020

.16

3.39

Yea

rssi

nce

mar

riag

e16

.45

6.44

16.3

16.

4816

.87

6.28

Edu

catio

nin

year

s3.

993.

543.

903.

584.

263.

42M

uslim

0.02

0.13

0.02

0.14

0.01

0.11

Chr

istia

n0.

960.

190.

960.

190.

970.

17H

asch

ildre

nag

ed5

and

unde

r0.

760.

430.

770.

420.

730.

44Ty

peof

resi

denc

e:R

ural

0.84

0.37

0.83

0.38

0.86

0.35

Hou

seho

ldha

sa

radi

o0.

630.

480.

620.

480.

640.

48H

ouse

hold

’sm

ain

floor

mat

eria

lis

cem

ent

0.17

0.37

0.17

0.38

0.16

0.37

Hou

seho

ldha

sel

ectr

icity

0.12

0.33

0.12

0.32

0.13

0.34

Hou

seho

ldw

ealth

isab

ove

the

med

ian

0.51

0.50

0.51

0.50

0.49

0.50

Obs

erva

tions

1165

287

9928

53

Not

e:Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

“Exp

osed

toth

em

ill”

repr

esen

tsbe

ing

inth

eca

tchm

ent

area

sof

am

ill.C

atch

men

tare

ara

dius

is4

km.

61

Tabl

eA

.3:S

umm

ary

Stat

istic

sfo

rWom

enba

sed

onTr

eatm

entS

tatu

s:20

05D

HS

Wom

en’s

Rec

ode

(Bef

ore

Rap

idE

xpan

sion

/Bas

elin

e)

Nev

erTr

eate

dSe

ctor

Lev

elB

efor

eTr

eatm

ent

Sect

orL

evel

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esW

orke

din

the

past

12m

onth

s0.

840.

370.

830.

38W

orke

dfo

rcas

hin

the

past

12m

onth

s0.

180.

390.

200.

40E

xper

ienc

eddo

mes

ticvi

olen

cein

the

past

12m

onth

s0.

220.

420.

220.

41

Pane

lB:C

ontr

ols

Hus

band

lives

inth

eho

use

0.89

0.31

0.86

0.35

Hus

band

’sag

e41

.24

9.74

41.5

49.

33H

usba

nd’s

Occ

upat

ion:

Agr

icul

tura

l0.

720.

450.

820.

39H

usba

nd’s

educ

atio

nin

year

s4.

163.

843.

773.

39O

ccup

atio

n:A

gric

ultu

ral

0.74

0.44

0.77

0.42

Mar

itals

tatu

s:M

arri

ed0.

620.

490.

700.

46M

onog

amy

(No

othe

rwiv

es)

0.86

0.34

0.88

0.32

Num

bero

funi

ons:

One

0.83

0.37

0.82

0.38

Age

atfir

stm

arri

age

19.6

73.

3319

.92

3.35

Yea

rssi

nce

mar

riag

e15

.08

6.85

15.4

07.

08E

duca

tion

inye

ars

3.56

3.59

3.64

3.21

Mus

lim0.

020.

140.

020.

14C

hris

tian

0.96

0.20

0.97

0.17

Has

child

ren

aged

5an

dun

der

0.85

0.36

0.83

0.38

Type

ofre

side

nce:

Rur

al0.

810.

390.

850.

36H

ouse

hold

has

ara

dio

0.53

0.50

0.53

0.50

Hou

seho

ld’s

mai

nflo

orm

ater

iali

sce

men

t0.

140.

350.

110.

31H

ouse

hold

has

elec

tric

ity0.

060.

240.

030.

18H

ouse

hold

wea

lthis

abov

eth

em

edia

n0.

510.

500.

470.

50

Obs

erva

tions

2566

905

Not

e:Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

“Tre

atm

ent”

repr

esen

tsa

mill

open

ing.

Sinc

ea

mill

serv

eson

lyto

itsca

tchm

enta

rea,

afte

rtre

atm

enti

sat

the

catc

hmen

tare

ale

velr

athe

rtha

nth

ese

ctor

leve

l.C

atch

men

tare

ara

dius

is4

km.

62

Tabl

eA

.4:S

umm

ary

Stat

istic

sfo

rWom

enba

sed

onTr

eatm

entS

tatu

s:20

05,2

010,

2014

DH

SW

omen

’sR

ecod

e

Nev

erTr

eate

dSe

ctor

Lev

elB

efor

eTr

eatm

ent

Sect

orL

evel

Aft

erTr

eatm

ent

Cat

chm

entA

rea

Lev

el

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esW

orke

din

the

past

12m

onth

s0.

880.

320.

840.

370.

880.

32W

orke

dfo

rcas

hin

the

past

12m

onth

s0.

370.

480.

260.

440.

470.

50E

xper

ienc

eddo

mes

ticvi

olen

cein

the

past

12m

onth

s0.

340.

470.

260.

440.

370.

48

Pane

lB:C

ontr

ols

Hus

band

lives

inth

eho

use

0.89

0.32

0.86

0.35

0.87

0.34

Hus

band

’sag

e42

.13

9.89

41.6

59.

3343

.11

9.72

Hus

band

’sO

ccup

atio

n:A

gric

ultu

ral

0.69

0.46

0.82

0.39

0.73

0.44

Hus

band

’sed

ucat

ion

inye

ars

4.43

3.90

3.92

3.38

4.32

3.60

Occ

upat

ion:

Agr

icul

tura

l0.

750.

430.

780.

410.

790.

41M

arita

lsta

tus:

Mar

ried

0.70

0.46

0.72

0.45

0.80

0.40

Mon

ogam

y(N

oot

herw

ives

)0.

880.

330.

880.

330.

920.

27N

umbe

rofu

nion

s:O

ne0.

840.

370.

830.

380.

850.

36A

geat

first

mar

riag

e19

.69

3.31

19.9

03.

2920

.16

3.39

Yea

rssi

nce

mar

riag

e16

.28

6.42

15.5

67.

0016

.87

6.28

Edu

catio

nin

year

s3.

943.

663.

683.

214.

263.

42M

uslim

0.02

0.14

0.02

0.14

0.01

0.11

Chr

istia

n0.

960.

190.

970.

180.

970.

17H

asch

ildre

nag

ed5

and

unde

r0.

770.

420.

820.

380.

730.

44Ty

peof

resi

denc

e:R

ural

0.82

0.38

0.86

0.34

0.86

0.35

Hou

seho

ldha

sa

radi

o0.

630.

480.

570.

500.

640.

48H

ouse

hold

’sm

ain

floor

mat

eria

lis

cem

ent

0.18

0.39

0.12

0.32

0.16

0.37

Hou

seho

ldha

sel

ectr

icity

0.13

0.34

0.04

0.19

0.13

0.34

Hou

seho

ldw

ealth

isab

ove

the

med

ian

0.52

0.50

0.49

0.50

0.49

0.50

Obs

erva

tions

7194

1041

2853

Not

e:Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

“Tre

atm

ent”

repr

esen

tsa

mill

open

ing.

Sinc

ea

mill

serv

eson

lyto

itsca

tchm

enta

rea,

afte

rtre

atm

enti

sat

the

catc

hmen

tare

ale

velr

athe

rtha

nth

ese

ctor

leve

l.C

atch

men

tare

ara

dius

is4

km.

63

Tabl

eA

.5:S

umm

ary

Stat

istic

sfo

rHus

band

s:D

HS

Cou

ple’

sR

ecod

e

All

Not

Exp

osed

toa

mill

Exp

osed

toa

mill

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esW

orke

din

the

past

12m

onth

s0.

870.

330.

870.

340.

890.

31W

orke

dfo

rcas

hin

the

past

12m

onth

s0.

800.

400.

790.

410.

820.

38

Pane

lB:C

ontr

ols

Hus

band

lives

inth

eho

use

1.00

0.07

1.00

0.07

0.99

0.08

Wif

e’s

Age

36.2

06.

8535

.96

6.87

36.9

76.

70W

ife’

sO

ccup

atio

n:A

gric

ultu

ral

0.76

0.43

0.75

0.43

0.79

0.41

Wif

e’s

Edu

catio

nin

year

s3.

913.

443.

843.

474.

143.

32O

ccup

atio

n:A

gric

ultu

ral

0.62

0.49

0.61

0.49

0.66

0.48

Mar

itals

tatu

s:M

arri

ed0.

770.

420.

740.

440.

840.

37M

onog

amy

(No

othe

rwiv

es)

0.95

0.22

0.94

0.23

0.96

0.20

Num

bero

funi

ons:

One

0.76

0.43

0.76

0.43

0.77

0.42

Age

atfir

stm

arri

age

23.7

14.

4223

.58

4.42

24.1

24.

41Y

ears

sinc

em

arri

age

17.1

67.

2517

.03

7.29

17.5

97.

09E

duca

tion

inye

ars

4.33

3.69

4.31

3.71

4.39

3.61

Mus

lim0.

020.

140.

020.

150.

010.

12C

hris

tian

0.95

0.21

0.95

0.22

0.96

0.20

Has

child

ren

aged

5an

dun

der

0.78

0.42

0.78

0.41

0.76

0.43

Type

ofre

side

nce:

Rur

al0.

840.

370.

830.

370.

870.

33H

ouse

hold

has

ara

dio

0.65

0.48

0.65

0.48

0.66

0.47

Hou

seho

ld’s

mai

nflo

orm

ater

iali

sce

men

t0.

160.

370.

160.

370.

160.

37H

ouse

hold

has

elec

tric

ity0.

120.

320.

110.

320.

130.

34H

ouse

hold

wea

lthis

abov

eth

em

edia

n0.

530.

500.

530.

500.

530.

50

Obs

erva

tions

4816

3657

1159

Not

e:Sa

mpl

eco

nsis

tsof

the

part

ners

ofw

omen

inTa

ble

A.5

.“E

xpos

edto

the

mill

”re

pres

ents

bein

gin

the

catc

hmen

tare

aof

am

ill.C

atch

men

tar

eara

dius

is4

km.

64

Tabl

eA

.6:S

umm

ary

Stat

istic

sfo

rHus

band

sba

sed

onTr

eatm

entS

tatu

s:D

HS

Cou

ple’

sR

ecod

e

Nev

erTr

eate

dSe

ctor

Lev

elB

efor

eTr

eatm

ent

Sect

orL

evel

Aft

erTr

eatm

ent

Cat

chm

entA

rea

Lev

el

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esW

orke

din

the

past

12m

onth

s0.

880.

330.

730.

440.

890.

31W

orke

dfo

rcas

hin

the

past

12m

onth

s0.

820.

390.

660.

470.

820.

38

Pane

lB:C

ontr

ols

Hus

band

lives

inth

eho

use

1.00

0.06

1.00

0.06

0.99

0.08

Wif

e’s

Age

35.9

76.

8135

.30

7.11

36.9

76.

70W

ife’

sO

ccup

atio

n:A

gric

ultu

ral

0.74

0.44

0.75

0.43

0.79

0.41

Wif

e’s

Edu

catio

nin

year

s3.

913.

603.

823.

144.

143.

32O

ccup

atio

n:A

gric

ultu

ral

0.59

0.49

0.57

0.49

0.66

0.48

Mar

itals

tatu

s:M

arri

ed0.

740.

440.

770.

420.

840.

37M

onog

amy

(No

othe

rwiv

es)

0.94

0.23

0.95

0.22

0.96

0.20

Num

bero

funi

ons:

One

0.77

0.42

0.75

0.43

0.77

0.42

Age

atfir

stm

arri

age

23.6

04.

4823

.98

4.47

24.1

24.

41Y

ears

sinc

em

arri

age

17.0

57.

1516

.11

7.70

17.5

97.

09E

duca

tion

inye

ars

4.47

3.85

3.97

3.34

4.39

3.61

Mus

lim0.

030.

160.

010.

110.

010.

12C

hris

tian

0.95

0.22

0.95

0.21

0.96

0.20

Has

child

ren

aged

5an

dun

der

0.78

0.42

0.84

0.37

0.76

0.43

Type

ofre

side

nce:

Rur

al0.

820.

380.

840.

360.

870.

33H

ouse

hold

has

ara

dio

0.65

0.48

0.61

0.49

0.66

0.47

Hou

seho

ld’s

mai

nflo

orm

ater

iali

sce

men

t0.

180.

380.

120.

330.

160.

37H

ouse

hold

has

elec

tric

ity0.

140.

350.

040.

200.

130.

34H

ouse

hold

wea

lthis

abov

eth

em

edia

n0.

540.

500.

500.

500.

530.

50

Obs

erva

tions

3219

612

1159

Not

e:Sa

mpl

eco

nsis

tsof

the

part

ners

ofw

omen

inTa

ble

A.5

.“Tr

eatm

ent”

repr

esen

tsa

mill

open

ing.

Sinc

ea

mill

serv

eson

lyto

itsca

tchm

enta

rea,

afte

rtre

atm

enti

sat

the

catc

hmen

tare

ale

velr

athe

rtha

nth

ese

ctor

leve

l.C

atch

men

tare

ara

dius

is4

km.

65

Tabl

eA

.7:S

umm

ary

Stat

istic

sfo

rHos

pita

ls:H

MIS

All

Not

Exp

osed

toa

mill

Exp

osed

toa

mill

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esH

asa

GB

VPa

tient

:Wom

enag

edol

dert

han

180.

790.

410.

850.

360.

700.

46H

asa

GB

VPa

tient

:Wom

enag

ed10

-18

yrs

0.84

0.36

0.89

0.31

0.77

0.42

Has

aG

BV

Patie

nt:M

enag

edol

dert

han

180.

240.

430.

270.

450.

190.

39H

asa

GB

VPa

tient

:Men

aged

10-1

8yr

s0.

130.

340.

150.

360.

110.

31D

eath

due

toG

BV

:Wom

enag

edol

dert

han

180.

020.

160.

020.

150.

030.

17D

eath

due

toG

BV

:Wom

enag

ed10

-18

yrs

0.01

0.08

0.01

0.08

0.01

0.08

Dea

thdu

eto

GB

V:M

enag

edol

dert

han

180.

020.

130.

020.

120.

020.

15D

eath

due

toG

BV

:Men

aged

10-1

8yr

s0.

010.

070.

000.

060.

010.

10

Pane

lB:B

reak

dow

nof

depe

nden

tvar

iabl

esin

num

bers

Patie

nts

w.p

hysi

calG

BV

sym

ptom

s:W

omen

aged

olde

rtha

n18

2.52

4.33

3.07

5.07

1.56

2.19

Patie

nts

w.s

exua

lGB

Vsy

mpt

oms:

Wom

enag

edol

dert

han

181.

702.

372.

022.

631.

161.

70Pa

tient

sw

.phy

sica

lGB

Vsy

mpt

oms.

:Wom

enag

ed10

-18

yrs

0.24

0.80

0.29

0.92

0.16

0.49

Patie

nts

w.s

exua

lGB

Vsy

mpt

oms:

Wom

enag

ed10

-18

yrs

5.07

5.24

5.58

5.49

4.36

4.67

Pane

lC:C

ontr

ols

Patie

ntis

refe

rred

toth

eho

spita

lby

polic

e0.

830.

380.

860.

350.

770.

42Pa

tient

isre

ferr

edto

the

hosp

italb

yco

mm

unity

heal

thw

orke

r0.

280.

450.

330.

470.

200.

40Pa

tient

isre

ferr

edfo

rcar

eto

high

erle

velh

ealth

faci

lity

0.08

0.28

0.09

0.28

0.08

0.27

Obs

erva

tions

1499

966

497

Not

e:G

BV

stan

dsfo

rgen

derb

ased

viol

ence

.Var

iabl

esin

Pane

lAan

dC

are

dum

my

vari

able

s.Pa

nelB

repo

rts

the

mon

thly

hosp

italiz

atio

nsfo

rG

BV.

Var

iabl

esre

port

edin

Pane

lBar

eus

edto

cons

truc

tthe

“has

aG

BV

patie

nt”

dum

my

vari

able

fora

spec

ific

age

grou

p.“P

atie

nt”

inPa

nelC

isa

gend

erba

sed

viol

ence

patie

nt.S

ampl

eco

nsis

tsof

hosp

itals

.HM

ISis

apa

neld

ata.

The

rear

e42

hosp

itals

whi

char

eob

serv

edfo

reve

rym

onth

for3

year

s.T

hus,

the

num

bero

fobs

erva

tions

isa

prod

ucto

fthe

num

bero

fhos

pita

ls,1

2an

d3.

“Exp

osed

toth

em

ill”

repr

esen

tsbe

ing

inth

eca

tchm

ent

area

ofa

mill

.Cat

chm

enta

rea

radi

usis

4km

.

66

Tabl

eA

.8:S

umm

ary

Stat

istic

sfo

rHos

pita

ls(n

onG

ende

rBas

edV

iole

nce

Hos

pita

lizat

ions

):H

MIS

All

Not

Exp

osed

toa

mill

Exp

osed

toa

mill

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:M

ain

depe

nden

tvar

iabl

esL

ogH

ospi

taliz

atio

ns:B

one

and

Jone

Dis

orde

rsot

hert

han

Frac

ture

s3.

101.

353.

161.

413.

011.

25

Pane

lB:D

epen

dent

vari

able

sin

leve

lsH

ospi

taliz

atio

ns:B

one

and

Jone

Dis

orde

rsot

hert

han

Frac

ture

s46

.35

86.9

552

.96

104.

5135

.51

42.9

0

Pane

lC:C

ontr

ols

Patie

ntis

refe

rred

toth

eho

spita

lby

polic

e0.

840.

370.

860.

340.

780.

42Pa

tient

isre

ferr

edto

the

hosp

italb

yco

mm

unity

heal

thw

orke

r0.

290.

450.

340.

470.

200.

40Pa

tient

isre

ferr

edfo

rcar

eto

high

erle

velh

ealth

faci

lity

0.08

0.27

0.09

0.28

0.07

0.25

Obs

erva

tions

3505

2279

1130

Not

e:G

BV

stan

dsfo

rgen

derb

ased

viol

ence

.Var

iabl

esin

Pane

lAan

dC

are

dum

my

vari

able

s.Pa

nelB

repo

rts

the

mon

thly

hosp

italiz

atio

nsfo

rbo

nean

djo

intd

isor

ders

othe

rtha

nfr

actu

res.

Sam

ple

cons

ists

ofho

spita

ls.H

MIS

isa

pane

ldat

a.T

here

are

42ho

spita

lsw

hich

are

obse

rved

for

ever

ym

onth

for3

year

s.T

hus,

the

num

bero

fobs

erva

tions

isa

prod

ucto

fthe

num

bero

fhos

pita

ls,1

2an

d3.

“Exp

osed

toth

em

ill”

repr

esen

tsbe

ing

inth

eca

tchm

enta

rea

ofa

mill

.Cat

chm

enta

rea

radi

usis

4km

.

67

Tabl

eA

.9:S

umm

ary

Stat

istic

sfo

rCou

ples

:EIC

V

All

Not

Exp

osed

toa

mill

Exp

osed

toa

mill

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:L

abor

inco

me

vari

able

sW

omen

’sla

stda

ilyea

rnin

gs,i

nR

wan

dan

fran

cs46

4.56

3064

.42

385.

0026

52.9

047

6.38

3120

.82

Hus

band

s’la

stda

ilyea

rnin

gs,i

nR

wan

dan

fran

cs11

86.6

553

18.5

710

21.1

454

71.9

912

11.2

352

95.1

8

Pane

lB:C

ontr

ols

Hus

band

isno

tabs

enti

nth

epa

st12

mon

ths

0.78

0.42

0.80

0.40

0.77

0.42

Hus

band

’sag

e40

.18

8.23

40.2

88.

1140

.16

8.25

Hus

band

’sO

ccup

atio

n:A

gric

ultu

ral

0.26

0.44

0.25

0.43

0.27

0.44

Hus

band

’sE

duca

tion:

No

educ

atio

n0.

110.

310.

120.

330.

110.

31H

usba

nd’s

Edu

catio

n:In

com

plet

epr

imar

y0.

560.

500.

520.

500.

560.

50H

usba

nd’s

Edu

catio

n:Pr

imar

yco

mpl

eted

0.26

0.44

0.27

0.44

0.26

0.44

Hus

band

’sE

duca

tion:

Mor

eth

anpr

imar

y0.

160.

370.

160.

370.

160.

36H

ouse

hold

size

5.78

1.95

6.05

1.99

5.74

1.94

Occ

upat

ion:

Agr

icul

tura

l0.

790.

400.

820.

380.

790.

41M

arita

lsta

tus:

Mar

ried

mon

ogam

ous

0.83

0.37

0.77

0.42

0.84

0.36

Mar

itals

tatu

s:M

arri

edpo

lyga

mou

s0.

040.

190.

050.

220.

030.

18E

duca

tion:

No

educ

atio

n0.

120.

320.

180.

390.

110.

31E

duca

tion:

Inco

mpl

ete

prim

ary

0.57

0.49

0.51

0.50

0.58

0.49

Edu

catio

n:Pr

imar

yco

mpl

eted

0.24

0.43

0.18

0.39

0.25

0.43

Edu

catio

n:M

ore

than

prim

ary

0.13

0.34

0.12

0.32

0.13

0.34

Chi

ldre

nin

the

HH

aged

5an

dun

der

0.79

0.41

0.82

0.39

0.79

0.41

Type

ofre

side

nce:

Rur

al0.

830.

380.

820.

380.

830.

37H

ouse

hold

’sm

ain

floor

mat

eria

lis

cem

ent

0.19

0.39

0.18

0.38

0.20

0.40

Hou

seho

ldha

sel

ectr

icity

0.15

0.36

0.12

0.33

0.15

0.36

Obs

erva

tions

1524

419

7113

273

Not

e:Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

pres

umab

lym

arri

edbe

fore

the

expa

nsio

nof

the

mill

s.Y

earo

fmar

riag

eis

notc

olle

cted

inE

ICV.

Inor

dert

oco

nstr

uctt

hesa

mpl

eof

wom

enw

hopa

rtne

red

befo

reth

eex

pans

ion

ofth

em

ills

(as

inD

HS)

,Ito

okth

eco

hort

sof

wom

enIo

bser

vein

DH

S.E

ICV

isno

tgeo

code

d.T

hus,

“Exp

osed

toth

em

ill”

repr

esen

tsha

ving

am

illin

the

dist

rict

.Dis

tric

tis

a80

0km

2ar

ea.

68

Tabl

eA

.10:

Sum

mar

ySt

atis

tics

forC

oupl

esba

sed

onTr

eatm

entS

tatu

s:E

ICV

Nev

erTr

eate

dD

istr

ictL

evel

Bef

ore

Trea

tmen

tD

istr

ictL

evel

Aft

erTr

eatm

ent

Dis

tric

tLev

el

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Mea

nSt

d.D

ev.

Pane

lA:L

abor

inco

me

vari

able

sW

omen

’sla

stda

ilyea

rnin

gs,i

nR

wan

dan

fran

cs54

5.79

3649

.95

330.

4219

30.7

947

5.45

3119

.86

Hus

band

s’la

stda

ilyea

rnin

gs,i

nR

wan

dan

fran

cs14

36.6

276

21.9

776

1.91

2128

.73

1216

.18

5314

.34

Pane

lB:C

ontr

ols

Hus

band

isno

tabs

enti

nth

epa

st12

mon

ths

0.78

0.42

0.83

0.37

0.77

0.42

Hus

band

’sag

e39

.65

7.97

40.8

58.

1340

.16

8.25

Hus

band

’sO

ccup

atio

n:A

gric

ultu

ral

0.28

0.45

0.22

0.41

0.26

0.44

Hus

band

’sE

duca

tion:

No

educ

atio

n0.

100.

290.

180.

380.

110.

31H

usba

nd’s

Edu

catio

n:In

com

plet

epr

imar

y0.

590.

490.

480.

500.

560.

50H

usba

nd’s

Edu

catio

n:Pr

imar

yco

mpl

eted

0.27

0.44

0.25

0.43

0.26

0.44

Hus

band

’sE

duca

tion:

Mor

eth

anpr

imar

y0.

150.

360.

180.

380.

160.

36H

ouse

hold

size

6.00

1.86

6.04

2.10

5.74

1.94

Occ

upat

ion:

Agr

icul

tura

l0.

850.

360.

760.

430.

790.

41M

arita

lsta

tus:

Mar

ried

mon

ogam

ous

0.83

0.37

0.71

0.45

0.84

0.36

Mar

itals

tatu

s:M

arri

edpo

lyga

mou

s0.

050.

210.

060.

230.

030.

18E

duca

tion:

No

educ

atio

n0.

150.

360.

230.

420.

110.

31E

duca

tion:

Inco

mpl

ete

prim

ary

0.58

0.49

0.46

0.50

0.58

0.49

Edu

catio

n:Pr

imar

yco

mpl

eted

0.16

0.37

0.21

0.41

0.25

0.43

Edu

catio

n:M

ore

than

prim

ary

0.11

0.31

0.15

0.35

0.13

0.34

Chi

ldre

nin

the

HH

aged

5an

dun

der

0.78

0.41

0.83

0.38

0.79

0.41

Type

ofre

side

nce:

Rur

al0.

860.

340.

750.

430.

830.

37H

ouse

hold

’sm

ain

floor

mat

eria

lis

cem

ent

0.13

0.34

0.23

0.42

0.20

0.40

Hou

seho

ldha

sel

ectr

icity

0.13

0.34

0.14

0.35

0.15

0.36

Obs

erva

tions

973

1593

1327

3

Not

e:Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

pres

umab

lym

arri

edbe

fore

the

expa

nsio

nof

the

mill

s.Y

earo

fmar

riag

eis

notc

olle

cted

inE

ICV.

Inor

dert

oco

nstr

uctt

hesa

mpl

eof

wom

enw

hopa

rtne

red

befo

reth

eex

pans

ion

ofth

em

ills

(as

inD

HS)

,Ito

okth

eco

hort

sof

wom

enIo

bser

vein

DH

S.E

ICV

isno

tgeo

code

d.T

hus,

“Tre

atm

ent”

repr

esen

tsa

mill

open

ing

inth

edi

stri

ct.D

istr

icti

sa

800

km2

area

.

69

Tabl

eA

.11:

Bal

ance

Che

ck:W

ithin

Dis

tric

tApp

roac

h

Hus

band

Wom

en

(1)

(2)

(3)

(4)

(5)

Occ

upat

ion:

Agr

icul

tura

lE

duca

tion

inY

ears

Occ

upat

ion:

Agr

icul

tura

lE

duca

tion

inY

ears

Civ

ilM

arri

age

Mill

-0.0

00.

10-0

.01

-0.0

8-0

.00

(0.0

2)(0

.15)

(0.0

2)(0

.14)

(0.0

2)

Obs

erva

tions

9823

9823

9823

9823

9823

Dep

ende

ntva

riab

lem

ean

0.71

4.34

0.76

4.02

0.73

Wom

enH

ouse

hold

(1)

(2)

(3)

(4)

(5)

Age

at:

Firs

tMar

riag

eR

elig

ion:

Chr

istia

nR

esid

ence

:R

ural

Cem

ent

Floo

rE

lect

rici

ty

Mill

0.01

-0.0

00.

02-0

.00

0.01

(0.0

6)(0

.01)

(0.0

3)(0

.01)

(0.0

1)

Obs

erva

tions

9823

9823

9823

9823

9823

Dep

ende

ntva

riab

lem

ean

19.9

00.

960.

830.

170.

12

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

pare

nthe

ses.

4km

catc

hmen

tare

ais

used

fort

hetr

eatm

entg

roup

.With

indi

stri

ctap

proa

chis

used

fort

heco

ntro

lgro

up.T

hees

timat

esar

eba

sed

onD

HS

data

and

estim

ated

with

the

mai

nsp

ecifi

catio

npr

esen

ted

inSe

ctio

n5.

2.1

***

p<.0

1,**

p<.0

5,*

p<.1

70

Tabl

eA

.12:

Bal

ance

Che

ck:D

onut

App

roac

h

Hus

band

Wom

en

(1)

(2)

(3)

(4)

(5)

Occ

upat

ion:

Agr

icul

tura

lE

duca

tion

inY

ears

Occ

upat

ion:

Agr

icul

tura

lE

duca

tion

inY

ears

Civ

ilM

arri

age

Mill

-0.0

1-0

.02

-0.0

1-0

.18

-0.0

1(0

.02)

(0.1

7)(0

.02)

(0.1

7)(0

.02)

Obs

erva

tions

4900

4900

4900

4900

4900

Dep

ende

ntva

riab

lem

ean

0.70

4.46

0.75

4.41

0.79

Wom

enH

ouse

hold

(1)

(2)

(3)

(4)

(5)

Age

at:

Firs

tMar

riag

eR

elig

ion:

Chr

istia

nR

esid

ence

:R

ural

Cem

ent

Floo

rE

lect

rici

ty

Mill

-0.0

10.

000.

04-0

.00

0.02

(0.0

7)(0

.01)

(0.0

3)(0

.01)

(0.0

1)

Obs

erva

tions

4900

4900

4900

4900

4900

Dep

ende

ntva

riab

lem

ean

20.1

50.

960.

810.

200.

16

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

pare

nthe

ses.

4km

catc

hmen

tare

ais

used

fort

hetr

eatm

entg

roup

.Don

utap

proa

chis

used

fort

heco

ntro

lgro

up.T

hees

timat

esar

eba

sed

onD

HS

data

and

estim

ated

with

the

mai

nsp

ecifi

catio

npr

esen

ted

inSe

ctio

n5.

2.1

***

p<.0

1,**

p<.0

5,*

p<.1

71

Tabl

eA

.13:

Eff

ecto

fMill

Exp

osur

eon

Wom

en’s

Occ

upat

ion:

With

inD

istr

ictA

ppro

ach

(1)

(2)

(3)

(4)

(5)

Man

ager

sSa

les

Agr

icul

tura

lSe

lf-E

mpl

oyed

Agr

icul

tura

lE

mpl

oyee

Man

ual

Skill

ed&

Uns

kille

d

Mill

0.00

-0.0

1-0

.01

-0.0

00.

02(0

.01)

(0.0

1)(0

.02)

(0.0

1)(0

.01)

Obs

erva

tions

8762

8762

8762

8762

8762

Dep

ende

ntva

riab

lem

ean

0.03

0.07

0.82

0.04

0.03

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

sis.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

All

depe

nden

tvar

iabl

esar

em

easu

red

fort

hepa

st12

mon

ths.

Cat

chm

enta

rea

radi

usis

4km

.With

indi

stri

ctap

proa

chis

used

fort

heco

ntro

lgro

up.*

**p<

.01,

**p<

.05,

*p<

.1

72

Tabl

eA

.14:

Eff

ecto

fMill

Exp

osur

eon

Hus

band

s’O

ccup

atio

n:W

ithin

Dis

tric

tApp

roac

h

(1)

(2)

(3)

(4)

(5)

Man

ager

sSa

les

Agr

icul

tura

lSe

lf-E

mpl

oyed

Agr

icul

tura

lE

mpl

oyee

Man

ual

Skill

ed&

Uns

kille

d

Mill

0.00

0.03

**-0

.04

0.00

0.01

(0.0

1)(0

.02)

(0.0

3)(0

.01)

(0.0

3)

Obs

erva

tions

3547

3547

3547

3547

3547

Dep

ende

ntva

riab

lem

ean

0.05

0.06

0.66

0.04

0.17

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

ses.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

ners

ofw

omen

inth

efir

stsp

ecifi

catio

n.A

llde

pend

entv

aria

bles

are

mea

sure

dfo

rthe

past

12m

onth

s.C

atch

men

tare

ara

dius

is4

km.W

ithin

dist

rict

appr

oach

isus

edfo

rthe

cont

rolg

roup

.***

p<.0

1,**

p<.0

5,*

p<.1

73

Tabl

eA

.15:

Eff

ecto

fMill

Exp

osur

eon

Wom

en’s

Occ

upat

ion

Don

utA

ppro

ach

(1)

(2)

(3)

(4)

(5)

Man

ager

sSa

les

Agr

icul

tura

lSe

lf-E

mpl

oyed

Agr

icul

tura

lE

mpl

oyee

Man

ual

Skill

ed&

Uns

kille

d

Mill

0.01

-0.0

1-0

.01

-0.0

10.

02(0

.01)

(0.0

1)(0

.02)

(0.0

1)(0

.02)

Obs

erva

tions

4628

4628

4628

4628

4628

Dep

ende

ntva

riab

lem

ean

0.03

0.08

0.80

0.03

0.04

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

ses.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

All

depe

nden

tvar

iabl

esar

em

easu

red

fort

hepa

st12

mon

ths.

Cat

chm

enta

rea

radi

usis

4km

.Don

utap

proa

chis

used

fort

heco

ntro

lgro

up.*

**p<

.01,

**p<

.05,

*p<

.1

74

Tabl

eA

.16:

Eff

ecto

fMill

Exp

osur

eon

Hus

band

s’O

ccup

atio

n:D

onut

App

roac

h

(1)

(2)

(3)

(4)

(5)

Man

ager

sSa

les

Agr

icul

tura

lSe

lf-E

mpl

oyed

Agr

icul

tura

lE

mpl

oyee

Man

ual

Skill

ed&

Uns

kille

d

Mill

0.01

0.04

**-0

.05

-0.0

10.

01(0

.01)

(0.0

2)(0

.04)

(0.0

2)(0

.03)

Obs

erva

tions

1958

1958

1958

1958

1958

Dep

ende

ntva

riab

lem

ean

0.05

0.06

0.67

0.04

0.16

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

ses.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

ners

ofw

omen

inth

efir

stsp

ecifi

catio

n.A

llde

pend

entv

aria

bles

are

mea

sure

dfo

rthe

past

12m

onth

s.C

atch

men

tare

ara

dius

is4

km.D

onut

appr

oach

isus

edfo

rthe

cont

rolg

roup

.***

p<.0

1,**

p<.0

5,*

p<.1

75

Table A.17: Effect of a Mill Opening on Women’s and Their Husbands’ Log of Last Daily Earningsusing Inverse Hypberbolic Sin Transformation for the Dependent Variables

All Sample Occupation: Agriculture

(1) (2) (3) (4)Woman: Log of

Last DailyEarnings


Woman: Log ofLast DailyEarnings


Log of Mills percapita in the District 2.778*** 2.539*** 2.236*** 2.786***

(0.19) (0.24) (0.17) (0.15)


Note: Robust standard errors clustered at the district level are in paranthesis. All estimations includeindividual controls, cohort fixed effects, district-by-year fixed effects, linear time trends interacted withbaseline district level characteristics. Since EICV is not geocoded, the mill variable is log of mills percapita in a district in a given year In Columns 3-4, the sample consists of women and their husbands whoreported their occupation as agricultural. *** p<.01, ** p<.05, * p<.1

Table A.18: Effect of a Mill Opening on Women-to-Husband Log of Last Daily Earnings Ratiousing Heckman Selection Model

All Sample

(1)Woman: Log of Last Daily Earnings

Log of Last Daily EarningsLog of Mills per

capita in the District 1.813***

(0.18)

Observations 5218

Note: Robust standard errors clustered at the district level are in parentheses. Estimates show the secondstage results. In both specifications, total number of children under the age 5 is used as the identifyingvariable and used in the first stage, selection equation. It affects women’s labor market participation(reservation wage), but not the wages (wage offer). All estimations include individual controls, cohort fixedeffects, district-by-year fixed effects, linear time trends interacted with baseline district level characteristics.In the selection equation, province fixed effects are used to avoid collinearity. Since EICV is not geocoded,the mill variable is log of mills per capita in a district in a given year. In Columns 4-6, the sample consistsof women and their husbands who reported their occupation as agricultural. *** p<.01, ** p<.05, * p<.1

76

Table A.19: Effect of Mill Exposure on Household’s Log Education Expenditures for ElementarySchool-Aged Children

(1)Household’s Log Education Expenditures for Children Aged 6-13

Log of Mills percapita in the District 1.26*

(0.73)

Observations 1530

Note: Robust standard errors clustered at the district level are in parentheses. Robust standard errorsclustered at the district level are in parentheses. All estimations include individual controls, cohort fixedeffects, district-by-year fixed effects, linear time trends interacted with baseline district levelcharacteristics. Since EICV is not geocoded, the mill variable is log of total number of mills per capita in adistrict a given year. Sample consists of households that have elementary school aged children (aged 6-13).*** p<.01, ** p<.05, * p<.1

77

Table A.20: Effect of Mill Exposure on Hospitalizations for Gender Based Violence

Women Men

(1) (2) (3) (4)Older

than 18Youngerthan 18

Olderthan 18

Youngerthan 18

Mill x Month 1 0.05 -0.01 0.06 -0.03(0.07) (0.05) (0.05) (0.06)

Mill x Month 2 0.00 0.00 0.00 0.00(.) (.) (.) (.)

Mill x Month 3 -0.04 -0.12 -0.10 -0.05(0.06) (0.10) (0.06) (0.08)

Mill x Month 4 -0.02 0.02 -0.08 0.01(0.08) (0.06) (0.07) (0.09)

Mill x Month 5 -0.01 -0.02 0.01 0.03(0.08) (0.11) (0.07) (0.10)

Mill x Month 6 -0.18** -0.07 -0.02 -0.01(0.07) (0.07) (0.10) (0.08)

Mill x Month 7 -0.14* 0.01 0.11 0.07(0.07) (0.10) (0.09) (0.12)

Mill x Month 8 0.06 -0.00 -0.04 0.04(0.07) (0.10) (0.09) (0.10)

Mill x Month 9 0.02 0.01 0.03 -0.05(0.07) (0.11) (0.09) (0.09)

Mill x Month 10 0.00 0.03 -0.06 -0.02(0.05) (0.09) (0.07) (0.11)

Mill x Month 11 -0.14 -0.06 -0.09 -0.02(0.11) (0.08) (0.07) (0.08)

Mill x Month 12 -0.07 0.07 -0.06 0.00(0.11) (0.09) (0.09) (0.07)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includehospital controls, hospital fixed effects, district fixed effects, province-by-year fixed effects, linear timetrends interacted with baseline district level characteristics. Omitted categeory is the hospitalizations inMonth 2, February. Catchment area radius is 4 km. *** p<.01, ** p<.05, * p<.1

78

Table A.21: Effect of Mill Exposure on Hospitalizations for Bone and Joint Disorders (Women)

Bone and Joint Disorders

(1)Women Older than 18

Mill x Month 1 0.14(0.18)

Mill x Month 2 0.00(.)


Mill x Month 4 -0.01(0.12)

Mill x Month 5 -0.08(0.16)

Mill x Month 6 -0.11(0.23)


Mill x Month 8 -0.02(0.19)

Mill x Month 9 -0.03(0.24)

Mill x Month 10 -0.04(0.22)

Mill x Month 11 -0.01(0.20)

Mill x Month 12 -0.03(0.19)

Observations 1078Dependent variable mean 20.37

Note: Fractures are excluded. Examples include osteoarthritis, gout, rheumatoid arthritis, lupus, bursitis.Dependent variable is in logs. Robust standard errors clustered at the sector level are in parentheses. Allestimations include hospital controls, hospital fixed effects, district fixed effects, province-by-year fixedeffects, linear time trends interacted with baseline district level characteristics. Omitted categeory is thehospitalizations in Month 2, February. Catchment area radius is 4 km. *** p<.01, ** p<.05, * p<.1

79

Table A.22: Hospitalizations for Gender Based Violence Among Women Older than 19 acrossDifferent Catchment Areas

Women Older than 19

(1) (2) (3)4 km 5 km 10 km

Mill x Month 1 0.05 0.05 0.11(0.07) (0.06) (0.10)

Mill x Month 2 0.00 0.00 0.00(.) (.) (.)

Mill x Month 3 -0.04 -0.04 0.08(0.06) (0.06) (0.10)

Mill x Month 4 -0.02 -0.01 0.06(0.08) (0.07) (0.11)

Mill x Month 5 -0.01 0.00 0.10(0.08) (0.07) (0.07)

Mill x Month 6 -0.18** -0.16** -0.12(0.07) (0.07) (0.08)

Mill x Month 7 -0.14* -0.13 0.01(0.07) (0.08) (0.10)

Mill x Month 8 0.06 0.05 -0.01(0.07) (0.07) (0.09)

Mill x Month 9 0.02 0.03 -0.02(0.07) (0.07) (0.07)

Mill x Month 10 0.00 0.00 -0.05(0.05) (0.04) (0.07)

Mill x Month 11 -0.14 -0.11 0.08(0.11) (0.10) (0.10)

Mill x Month 12 -0.07 -0.04 0.08(0.11) (0.10) (0.10)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includehospital controls, hospital fixed effects, district fixed effects, province-by-year fixed effects, linear timetrends interacted with baseline district level characteristics. Omitted category is the hospitalizations inMonth 2, February. Catchment area radius is 4 km. *** p<.01, ** p<.05, * p<.1

80

Table A.23: Effect of Mill Exposure on Women’s Employment and Type of Earnings

Work Variables

(1) (2)Work Cash

Mill x Event Time -3 0.02 0.05(0.03) (0.06)

Mill x Event Time -2 0.02 0.05(0.03) (0.07)

Mill x Event Time -1 0.00 0.00(.) (.)

Mill x Event Time 0 -0.01 0.10*(0.02) (0.06)

Mill x Event Time 1 0.00 0.06(0.03) (0.08)

Mill x Event Time 2 0.00 0.07(0.03) (0.08)

Mill x Event Time 3 0.00 0.06(0.02) (0.06)

Mill x Event Time 4 0.00 0.09**(0.02) (0.04)

Mill x Event Time 5 -0.03 0.17***(0.03) (0.06)

Mill x Event Time 6 0.01 0.10**(0.02) (0.04)

Observations 7827 8163Dependent variable mean 0.89 0.41

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All variables are measured for the past 12months. Ommitted categeory is Event Time -1, 1 year before a mill opening. Catchment area radius is 4km. *** p<.01, ** p<.05, * p<.1

81

Table A.24: Effect of Mill Exposure on Self-Reported Domestic Violence

(1)Domestic Violence

Mill x Event Time -3 0.01(0.07)

Mill x Event Time -2 -0.06(0.08)

Mill x Event Time -1 0.00(.)

Mill x Event Time 0 -0.16***(0.06)

Mill x Event Time 1 -0.17**(0.09)

Mill x Event Time 2 -0.12(0.09)

Mill x Event Time 3 -0.14**(0.06)

Mill x Event Time 4 -0.05(0.05)

Observations 3528Dependent variable mean 0.35

Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All variables are measured for the past 12months. Omitted category is Event Time -1, 1 year before a mill opening. Catchment area radius is 4 km.*** p<.01, ** p<.05, * p<.1

82

Table A.25: Effect of Mill Exposure on Women’s Employment and Type of Earnings across Dif-ferent Catchment Areas

4 km 5 km 10 km

(1) (2) (3) (4) (5) (6)Work Cash Work Cash Work Cash

Mill x Event Time -3 0.01 0.03 0.00 -0.03 -0.02 -0.03(0.03) (0.05) (0.02) (0.06) (0.02) (0.04)

Mill x Event Time -2 0.02 0.07 0.01 -0.00 -0.04 0.04(0.02) (0.05) (0.03) (0.04) (0.03) (0.05)

Mill x Event Time -1 0.00 0.00 0.00 0.00 0.00 0.00(.) (.) (.) (.) (.) (.)

Mill x Event Time 0 -0.01 0.04 -0.02 0.04 -0.03 0.04(0.02) (0.06) (0.02) (0.04) (0.02) (0.04)

Mill x Event Time 1 0.01 0.06 0.01 0.06 0.01 0.03(0.03) (0.08) (0.02) (0.06) (0.03) (0.06)

Mill x Event Time 2 -0.00 0.05 0.00 0.11* -0.01 0.01(0.03) (0.07) (0.03) (0.07) (0.02) (0.05)

Mill x Event Time 3 -0.01 0.07 -0.03 -0.01 -0.05* -0.02(0.02) (0.06) (0.02) (0.04) (0.03) (0.05)

Mill x Event Time 4 0.02 0.09** -0.03 -0.01 -0.01 0.06(0.02) (0.04) (0.02) (0.04) (0.02) (0.04)

Mill x Event Time 5 -0.04* 0.14** -0.05** 0.01 0.00 0.09**(0.02) (0.06) (0.02) (0.06) (0.02) (0.04)

Mill x Event Time 6 0.01 0.08** -0.01 -0.00 -0.01 -0.03(0.02) (0.04) (0.02) (0.04) (0.02) (0.04)


Note: Robust standard errors clustered at the sector level are in paranthesis. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All variables are measured for the past 12months. Omitted category is Event Time -1, 1 year before a mill opening. *** p<.01, ** p<.05, * p<.1

83

Table A.26: Effect of Mill Exposure on Self-Reported Domestic Violence across Different Catch-ment Areas

Domestic Violence

(1) (2) (3)4 km 5 km 10 km

Mill x Event Time -3 0.00 -0.00 0.00(0.07) (0.06) (0.07)

Mill x Event Time -2 -0.06 -0.09 0.01(0.08) (0.07) (0.09)

Mill x Event Time -1 0.00 0.00 0.00(.) (.) (.)

Mill x Event Time 0 -0.16** -0.09 -0.02(0.07) (0.06) (0.06)

Mill x Event Time 1 -0.14* -0.08 -0.03(0.08) (0.09) (0.08)

Mill x Event Time 2 -0.13 -0.04 -0.05(0.09) (0.08) (0.07)

Mill x Event Time 3 -0.14** -0.09 0.01(0.06) (0.08) (0.08)

Mill x Event Time 4 -0.05 -0.01 0.01(0.05) (0.05) (0.06)


Note: Robust standard errors clustered at the sector level are in parentheses. All estimations includeindividual controls, cohort fixed effects, year of marriage fixed effects, sector fixed effects, district-by-yearfixed effects, linear time trends interacted with baseline sector level characteristics. Sample consists ofpartnered women who married before the expansion of the mills. All variables are measured for the past 12months. Omitted category is Event Time -1, 1 year before a mill opening. *** p<.01, ** p<.05, * p<.1

84

Tabl

eA

.27:

Eff

ecto

fM

illE

xpos

ure

onW

omen

’sE

mpl

oym

ent,

Type

ofE

arni

ngs

and

Self

-Rep

orte

dD

omes

ticV

iole

nce

inth

ePa

st12

Mon

ths

forD

iffer

entC

atch

men

tAre

aSi

zes

4km

5km

10km

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Wor

kC

ash

Vio

lenc

eW

ork

Cas

hV

iole

nce

Wor

kC

ash

Vio

lenc

e

Mill

-0.0

00.

07**

*-0

.09*

**-0

.01

0.04

**-0

.07*

-0.0

1-0

.02

0.02

(0.0

1)(0

.02)

(0.0

3)(0

.01)

(0.0

2)(0

.04)

(0.0

1)(0

.03)

(0.0

4)

Obs

erva

tions

9823

8766

3583

9823

8766

3583

9823

8766

3583

Dep

ende

ntva

riab

lem

ean

0.88

0.40

0.35

0.88

0.40

0.35

0.88

0.40

0.35

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

pare

nthe

ses.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

All

depe

nden

tvar

iabl

esar

em

easu

red

fort

hepa

st12

mon

ths.

Cat

chm

enta

reas

are

cons

truc

ted

bybu

ffer

sar

ound

the

mill

sw

here

4,5

and

10km

repr

esen

tbuf

ferr

adiu

s.**

*p<

.01,

**p<

.05,

*p<

.1

85

Tabl

eA

.28:

Eff

ecto

fMill

Exp

osur

eon

Hou

seho

ldD

ecis

ion

Mak

ing:

Wom

anA

lone

orJo

intly

with

Hus

band

4km

5km

10km

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Lar

geH

HPu

rcha

ses

Ow

nH

ealth

Fam

ilyV

isit

Lar

geH

HPu

rcha

ses

Ow

nH

ealth

Fam

ilyV

isit

Lar

geH

HPu

rcha

ses

Ow

nH

ealth

Fam

ilyV

isit

Mill

0.05

**0.

020.

020.

04*

0.01

0.01

0.02

0.01

0.01

(0.0

2)(0

.02)

(0.0

2)(0

.02)

(0.0

2)(0

.02)

(0.0

2)(0

.02)

(0.0

2)

Obs

erva

tions

9823

9823

9823

9823

9823

9823

9823

9823

9823

Dep

ende

ntva

riab

lem

ean

0.69

0.73

0.81

0.69

0.73

0.81

0.69

0.73

0.81

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

pare

nthe

ses.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

All

depe

nden

tvar

iabl

esar

eco

ded

as1

ifth

ede

cisi

onis

take

nby

wom

enal

one

orjo

intly

with

thei

rhus

band

san

dze

root

herw

ise.

Oth

erw

ise

cate

gory

incl

udes

husb

and

only

orso

meo

neel

se.T

hesu

rvey

ques

tion

onho

useh

old

deci

sion

sch

ange

dsi

gnifi

cant

lyov

erth

eda

tacy

cles

.in

2005

,the

ques

tion

asks

“Who

has

the

final

say

onde

cisi

onx?

”A

fter

2005

,the

ques

tion

is“W

hous

ually

deci

des

onx”

.The

answ

erop

tion

“Res

pond

enta

ndot

her”

isal

sodi

scon

tinue

din

2010

and

2014

roun

ds.T

om

ake

aco

mpa

rabl

eva

riab

le,I

crea

teth

eva

riab

leba

sed

onw

heth

erth

ew

oman

had

asa

yin

the

deci

sion

,eith

eral

one

orjo

intly

.Res

ults

shou

ldbe

inte

rpre

ted

with

caut

ion.

Cat

hmen

tare

asar

eco

nstr

ucte

dby

buff

ers

arou

ndth

em

ills

whe

re4,

5an

d10

kmre

pres

entb

uffe

rrad

ius.

***

p<.0

1,**

p<.0

5,*

p<.1

86

Figu

reA

.1:F

AO

-GA

EZ

Cof

fee

Suita

bilit

yIn

dex

and

Exp

ansi

onof

Mill

sin

Rw

anda

Not

e:T

hem

aps

are

cons

truc

ted

byco

mbi

ning

data

onm

ills,

FAO

-GA

EZ

Cof

fee

Cof

fee

Suita

bilit

yIn

dex,

The

Rw

anda

nG

eoPo

rtal

spat

iald

ata

onse

ctor

boun

drie

s.

87

Figu

reA

.2:H

ospi

tals

inR

wan

daN

ote:

The

map

isco

nstr

ucte

dby

com

bini

ngH

MIS

data

with

the

GPS

coor

dina

tes

ofth

eho

spita

lspr

ovid

edby

the

Rw

anda

Min

istr

yof

Hea

lth.

88

Figure A.3: Dynamic Impact of a Mill Opening on Hospitalizations for Gender Based Violence(Women 10-18)


89

Figure A.4: Mean Hospitalizations for Gender Based Violence within a Year

Note: This figure shows the raw means for hospitalizations for women older than 18 within a year.

90

Figure A.5: Dynamic Impact of a Mill Opening on Hospitalizations for Gender Based Violence(Men)


91

Figure A.6: Impact of a Mill Opening on Women’s Employment, Type of Earnings and Self-Reported Domestic Violence in the Past 12 Months using de Chaisemartin and D’Haultfœuille(2020)

Note: Sample consists of partnered women who married before the expansion of the mills. All variablesare measured for the past 12 months. 4 km catchment area is used. Within district approach is used.

92

Figure A.7: Dynamic Impact of a Mill Opening on Women’s Employment, Type of Earnings andSelf-Reported Domestic Violence in the Past 12 Months using Sun and Abraham (2020)

Note: Sample consists of partnered women who married before the expansion of the mills. All variablesare measured for the past 12 months. 4 km catchment area is used. Within district approach is used.

93

Figu

reA

.8:D

ynam

icIm

pact

ofa

Mill

Ope

ning

onD

eath

sdu

eto

Gen

derB

ased

Vio

lenc

e(W

omen

and

Men

)

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

pare

nthe

ses.

All

estim

atio

nsin

clud

eho

spita

lcon

trol

s,ho

spita

lfixe

def

fect

s,di

stri

ctfix

edef

fect

s,pr

ovin

ce-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

dist

rict

leve

lcha

ract

eris

tics.

Cat

chm

enta

rea

radi

usis

4km

..**

*p<

.01,

**p<

.05,

*p<

.1

94

Figu

reA

.9:

Dyn

amic

Impa

ctof

aM

ill’s

Peri

odof

Ope

ratio

non

Hos

pita

lizat

ions

forG

ende

rBas

edV

iole

nce

(Wom

enA

ged

Old

erth

an19

)for

Diff

eren

tCat

chm

entA

reas

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

sis.

All

estim

atio

nsin

clud

eho

spita

lcon

trol

s,ho

spita

lfixe

def

fect

s,di

stri

ctfix

edef

fect

s,pr

ovin

ce-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

dist

rict

leve

lcha

ract

eris

tics.

4km

,5km

and

10km

buff

erra

diia

reus

edre

spec

tivel

yfo

rthe

catc

hmen

tare

a.**

*p<

.01,

**p<

.05,

*p<

.1

95

Figu

reA

.10:

Dyn

amic

Impa

ctof

aM

illO

peni

ngon

Wom

en’s

Em

ploy

men

t,Ty

peof

Ear

ning

san

dSe

lf-R

epor

ted

Dom

estic

Vio

lenc

ein

the

Past

12M

onth

sfo

rDiff

eren

tCat

chm

entA

reas

Not

e:R

obus

tsta

ndar

der

rors

clus

tere

dat

the

sect

orle

vela

rein

para

nthe

sis.

All

estim

atio

nsin

clud

ein

divi

dual

cont

rols

,coh

ortfi

xed

effe

cts,

year

ofm

arri

age

fixed

effe

cts,

sect

orfix

edef

fect

s,di

stri

ct-b

y-ye

arfix

edef

fect

s,lin

eart

ime

tren

dsin

tera

cted

with

base

line

sect

orle

velc

hara

cter

istic

s.Sa

mpl

eco

nsis

tsof

part

nere

dw

omen

who

mar

ried

befo

reth

eex

pans

ion

ofth

em

ills.

All

vari

able

sar

em

easu

red

fort

hepa

st12

mon

ths.

5km

and

10km

buff

erra

diia

reus

edre

spec

tivel

yfo

rthe

catc

hmen

tare

a.**

*p<

.01,

**p<

.05,

*p<

.1

96

B Theoretical Appendix

The probability of the husband choosing violence is

Fh

[(1− y)Ih− (1− x+ τ)Iw−α(Ih + Iw)+

τIw +α(Ih + Iw)

Fw[(1− x+ τ)Iw− (1− y)Ih + v]

]. (12)

B.1 Increase in Iw

∂Fh

∂ Iw= fw

[(1− y)Ih− (1− x+ τ)Iw−αh(Ih + Iw)+

τIw +α(Ih + Iw)

Fw[(1− x+ τ)Iw− (1− y)Ih + v]

]×

[[− (1− x+ τ)− ∂α(Ih+Iw)

∂ Iw

]Fw[(1− x+ τ)Iw− (1− y)Ih + v]2

Fw[(1− x+ τ)Iw− (1− y)Ih + v]2+

(τ + ∂α(Ih+Iw)∂ Iw

)Fw[(1− x+ τ)Iw− (1− y)Ih + v]− fw[(1− x+ τ)Iw− (1− y)Ih + v](1− x+ τ)[τIw +α(Iw, Ih)]

Fw[(1− x+ τ)Iw− (1− y)Ih + v]2

].

Assume Fw(.)> 0 and fw(.)> 0. Then the first line of the product is positive. The right handside of the product can be both negative and positive because Fw(.)> 0, fw(.)> 0, (1−x+τ)> 0,∂αh(Ih+Iw)

∂ Ih< 0, τIw > 0 and α(Ih + Iw) > 0. Thus the sign of the derivative is ambiguous. The

derivative is negative when

κ =[− (1− x+ τ)− ∂α(Ih + Iw)

∂ Iw

]Fw[(1− x+ τ)Iw− (1− y)Ih + v]2+

(τ +

∂α(Ih + Iw)

∂ Iw

)Fw[(1− x+ τ)Iw− (1− y)Ih + v]

− fw[(1− x+ τ)Iw− (1− y)Ih + v](1− x+ τ)[τIw +α(Iw + Ih)]< 0.

When Iw is large enough, and τ and | ∂αh(Ih+Iw)∂ Iw

| are small enough, κ < 0 and the derivative isnegative.

97

B.2 Increase in Ih

∂Fh

∂ Ih= fh

[(1− y)Ih− (1− x+ τ)Iw−α(Ih + Iw)+

τIw +α(Ih + Iw)

Fw[(1− x+ τ)Iw− (1− y)Ih + v]

]×

[[(1− y)− ∂αh(Ih+Iw)

∂ Ih

]Fw[(1− x+ τ)Iw− (1− y)Ih + v]2

Fw[(1− x+ τ)Iw− (1− y)Ih + v]2+

∂α(Ih+Iw)∂ Ih

Fw[(1− x+ τ)Iw− (1− y)Ih + v]+ fw[(1− x+ τ)Iw− (1− y)Ih + v](1− y)(τIw +α(Ih + Iw))

Fw[(1− x+ τ)Iw− (1− y)Ih + v]2

].

Assuming that Fw(.) and fh(.)> 0, the left hand side is positive. The right hand side of the productcan be both positive or negative because Fw(.)> 0, fw(.)> 0, (1− y)> 0, ∂αh(Ih+Iw)

∂ Ih< 0, τIw > 0

and αh(Ih + Iw)> 0. Thus the sign of the derivative is ambiguous. The derivative is negative when

φ =[(1− y)− ∂α(Ih + Iw)

∂ Ih

]Fw[(1− x+ τ)Iw− (1− y)Ih + v]2+

∂α(Ih + Iw)

∂ IhFw[(1−x+τ)Iw−(1−y)Ih+v]+ fw[(1−x+τ)Iw−(1−y)Ih+v](1−y)(τIw+αh(Ih+Iw))< 0

When ∂αh(Ih+Iw)∂ Ih

is small enough, φ < 0 and the derivative is negative.

98

C Data Appendix

C.1 Panel of Mills

First, I complement the mills data with spatial data by The Rwanda GeoPortal. The portal providesmaps of different geographical boundries like sector, district and province. There are 416 sectors,30 districts and 5 provinces in Rwanda. The mean area of each is approximately 58.4 km2, 810km2 and 5552 km2. According to 2002 Rwandan Population and Housing Census, average sectorpopulation is 1959 where average female population at the sector level is 1024. Average districtpopulation is 27698 where average female population at the district level is 14499.

The final panel of mills have information on a mill’s location at different levels (including theexact GPS coordinates), number of coffee trees in the sector the mill is located for different yearsand FAO-GAEZ coffee suitability score. The data also has information on mill characteristics andspatial features of the location of the mill including presence of water bodies, road network, areaetc. Figure 2 visualizes the finalized data. The years are selected based on the availability of thecoffee census and data on women’s outcome variables (2004, 2005, 2010 and 2014). The rapidexpansion of the mills between 2005 and 2010 are observed in the areas that had a high number ofcoffee trees in 1999.

Table A.1 provides summary statistics. 78% of the mills started to operate between 2005-2010,the rapid expansion period. 50% of them are owned by cooperatives where only 25% of them areowned by NGOs. The remaining are owned by entrepreneurs or private companies. Log numberof coffee trees increased over time in the sectors where mills are built. 65% of the mills are built inareas where FAO-GAEZ coffee suitability index is either moderate, medium or good. Accordingto Macchiavello and Morjaria (2020), the mean number of coffee farmers that sell a mill is 396(number of coffee farmers in the catchment area of a mill).41

C.2 Individuals and Household Level Data

Rwandan Demographic Health Surveys. DHS asks whether women are paid “cash only” or“cash and in-kind”. Field surveys suggest that mills provide clothing, in-kind benefits to womenworking in the mills beyond giving a daily wage. Thus, I create my cash variable based on womenworking for both cash and in-kind rather than cash only. According to the data, occupations thatpay cash-only wages for women are dominantly non-agricultural (managers in Kigali, sales etc.).Upon a mill opening, I find statistically significant changes in the probability of working for cash

41Macchiavello and Morjaria (2020) measures the catchment area with a 5 km radius buffer zone.

99

when I define the variable as women working for both cash and in kind. However, there is nochange in the probability of working for cash upon a mill opening when I define the variable aswomen working for cash only. This further confirms that working for “cash and in-kind” correctlycaptures the type of earnings of the women who work in the mills. I perform the same exercise forhusbands. There is no change in husbands’ working for cash behavior for both definitions whichconfirms that the expansion did not affect men’s earning type.

Tables A.2 and A.5 provide descriptive statistics for women and their husbands. 88% of womenworked in the past 12 months, where only 39% of them worked for cash. The share of workingwomen is the same across areas that are exposed and not exposed to a mill. However, 47% ofwomen are working for cash in mill exposed areas where this number is 37% for women in unex-posed areas. The dominant occupation for women is agricultural where women are self-employed(farmers). 79% and 76% of women are working in agriculture in exposed and unexposed areasrespectively. Agriculture is the dominant occupation for women’s husbands as well. 73% and 71%of the husbands are working in agriculture in exposed and unexposed areas respectively. In con-trast to women’s outcomes, for husbands, the share of working and working for cash in the past12 months are both similar across both areas. Tables A.3, A.4 and A.6 provide summary statisticsbased on treatment status (averages for never-treated sectors, sectors before a mill opening andinside the catchment areas) which are all in line with the averages in Tables A.2 and A.5.

The women’s questionnaire also has a domestic violence module which collects informationon family violence.42 DHS randomly selects one women per household for the module. Thus, thenumber of women who have information for domestic violence will always be less than the numberof women who answered the women’s questionnaire. The module asks partnered women whetherthey experience physical, sexual or emotional violence in the last 12 months by their partners.

Since emotional violence questions are not asked in the 2010/11 cycle, I use physical and sexualdomestic violence in my analysis. Physical domestic violence consists of being pushed, shaken,thrown something at, slapped, kicked, dragged, strangled, burnt and sexual domestic violenceconsists of physically forced into unwanted sex and perform sexual acts.

I also exploit variables related to household-decision making to analyze the relationship be-tween exposure to mills and women’s bargaining power within the household. DHS asks womenwhether they make certain decisions alone, jointly with their husband or the decisions are made forher by their husbands or someone else. Such decisions include decisions regarding large householdpurchases, women’s own health and women visiting their own family.

The surveys also collect GPS coordinates for every cluster of households. Using the GPS42The module is prepared with respect to WHO guidelines “Putting Women First: Ethical and Safety Recommen-

dations for Research on Domestic Violence against Women” World Health Organization, 2001.

100

coordinates, I spatially merge the DHS data with the mills data. One potential concern is thatDHS randomly displace the GPS coordinates to maintain confidentiality. Due to the random dis-placement, urban clusters contain a minimum of 0 and a maximum of 2 kilometers of error andrural clusters contain a minimum of 0 and a maximum of 5 kilometers of positional error. Thus,GPS displacement may leads to measurement error and can bias the results (Perez-Heydrich et al.,2013). In order to reduce distance measurement error, I follow Perez-Heydrich et al. (2013), whichsuggests using a buffer distance rather than a closest distance when using a distance measure. Allthe distances calculated in the paper are based on a buffer distance measure. Thus, exposure to millare constructed via creating buffers. Details on measuring exposure to mills are outlined at the endof the section.

DHS data allows me to observe the outcome variables for 2004, 2005, 2010 and 2014. The2005 cycle was collected in February-July 2005, the 2010/11 cycle was collected in September2010-March 2011 and the 2014/15 cycle was collected in November 2014-April 2015. Given thatthe harvest season runs from April to July and key variables captures individuals’ experience inthe past 12 months, 2005 data cycle contains information on both 2004 and 2005 harvest season.The 2010/11 cycle contains information on 2010 harvest season and the 2014/15 cycle providesinformation on 2014 harvest season. Thus, DHS data enables me to observe 4 harvest years intotal, 2004, 2005, 2010 and 2014.

Integrated Household Living Conditions Surveys. Table A.9 provides descriptive statistics forwomen and their husbands.

Since EICV does not provide information on the year of marriage, unfortunately, I cannotrestrict the sample to couples who married before the expansion. However, I include cohort fixedeffects in my specifications that can provide a control for the year of marriage. Moreover, I restrictthe sample to individuals who are in the same age groups with the DHS data.

C.3 Administrative Hospital Level Domestic Violence Data

District hospitals constitute the overwhelming majority of the hospitals (47 in total) in the coun-try.43 The remaining 5 hospitals are either referral and teaching hospitals. Referral hospitals arelocated either in Kigali (capital) or urban areas, have the latest technology and provide care underthe public national health insurance system if a district hospital or health clinic refers a currentpatient to visit a referral hospital. Thus, they do not constitute a first stop for a patient. Teachinghospitals focus on medical research.

43Given that there are in total 30 districts, some districts have more than one district hospital.

101

Figure A.2 visualizes the hospitals on a map. The district hospitals have gender based violencecenters in them which aim to combat gender based violence. These centers, also known as theIsange One Stop Centers (IOSCs), are state institutions which provide medical, police and legalhelp for gender based violence victims under one roof and are free of charge. By combining allservices under one roof, MOH, Rwanda National Police, the Ministry of Justice (MOJ) and theMinistry of Gender and Family Promotion (MIGEPROF) aim to ease the reporting process of gen-der based violence victims. Sviatschi and Trako (2021) finds that such centers in Peru reducedgender-based violence and increased human capital investments in children, raised children’s en-rollment, attendance, and test scores.

Table A.7 provides descriptive statistics for the hospitals. 79% of the hospitals have hosted atleast a domestic violence patient (GBV patient older than 18) in a given month. The share is 70%for the hospitals who are exposed to the mills where it is 85% for the ones who are not exposed.Mean number of patients with physical domestic violence and sexual domestic violence symptomsin a month in a hospital is 2.52 and 1.71 respectively. The numbers are 1.56 and 1.16 for thehospitals who are exposed to the mills. For the hospitals who are not exposed to the mills, themean number of patients with physical domestic violence and sexual domestic violence symptomsin a month are 3.07 and 2.02 respectively.

To perform a placebo test, I requested additional data from the Rwandan Ministry Health, hos-pitalizations for non-domestic violence diseases. I receive bone and joint disorders other thanfractures. Table A.8 provides descriptive statistics for the hospitals. Osteoarthritis, gout, rheuma-toid arthritis, lupus, bursitis are examples of such diseases. Injuries and broken bones are excludedwhich rules out the concern that this group of hospitalizations also capture domestic violence.Mean of hospitalizations is 46 in a month.

102

paid work for women and domestic violence: evidence from

Documents